Categories
Biomedical Research

Promise and Peril: Machine Learning in Modern Cancer Treatment

A discussion of both the promises and perils of artificial intelligence in cervical cancer, brain cancer, and lung cancer treatment. A fast-growing field that promises many scientific breakthroughs in future.

By Sia Shah

Published 4:03 PM EST, Mon June 7, 2021

Abstract

Machine learning in modern cancer treatment is a fast-growing field that promises to produce many scientific breakthroughs in the future. This article discusses both the promises and perils that come along with applying artificial intelligence to cancer treatment. With cervical cancer treatment, this growing technology can be used to assist doctors in cancer detection as well as to predict patient survival rates. In lung cancer treatment, artificial intelligence platforms are again used to make predictions for patient health in addition to analyzing images for a more accurate prognosis. Finally, machine learning is also able to predict the survival rate and metastasis for different forms of brain cancer and provide medical students with realistic surgical simulations on how to operate. However, while there are a multitude of promises for the future of AI in medicine, integrating new technologies into a previously established field does have disadvantages. The constant evolution of software and technology means that operators require constant training to be able to handle the tasks Furthermore, the lack of doctor-patient feedback can take a negative toll on patients’ mental health and privacy. The automation of various processes, comes at the cost of various jobs, of people who originally performed these tasks manually. Therefore, when implementing AI into the medical field it is important to acknowledge the great promises the technology has, but also to weigh the negative effects that may result from its application.

Introduction

One of the most important parts of practicing medicine is decision-making, a skill that relies heavily on judgment. Cancer treatment, or Oncology, is a medical speciality where decision-making is incredibly important because of the unpredictable responses to treatment and change in a patient’s condition. This is where artificial intelligence (AI) comes into play. It is a promising tool that can objectively interpret cancer images and predict a cancer patient’s outcome- essentially mimicking the cognitive functions of humans. Research has shown that AI has the potential to exceed human performance in certain areas of medicine. Multiple examples of useful areas within AI will be discussed in this paper including two main ones. The first is detection or determining which objects are located within the body by analyzing images. The second task is characterization, separating tumours into groups based on physical appearance (Bi et al., 2019). Both of these tasks are a crucial part of making clinical decisions.

Machine learning (ML) is a subset of AI (Fig.1) that has been widely used in current healthcare applications since it uses data to train computational systems without the need for explicit programming. These computer programs can learn and improve from experience, unlike traditional computer programs that require specific instruction at each step, which makes them incredibly useful in the field of science (Ahuja, 2019). Allowing machines to make predictions based on a pattern that they have recognized. With the use of ML, a computer can use previously labeled data or even the pattern found in the data, and make predictions about it. In particular, ML excels at finding indistinct patterns which are undetectable to humans, in larger sets of data. ML also enables an algorithm to perform a task such as making medical decisions or driving a car while also correcting its own mistakes. Deep learning is a subset of ML that uses structures similar to a brain neural network in order to identify patterns within large datasets. CNN’s or convolutional neural networks are other subsets of ML that will be discussed and are generally applied to classification as well as analysis of patient scans (Hashimoto, Rosman, Rus, & Meireles, 2018).

In the future, AI analysis has the potential to work its way into all parts of patient care. Before surgery, it can help track the activity of a patient and access electronic health records. During surgery, it could assist the surgeon in making quick decisions based on the patient’s vital signs. After surgery, it can continue to collect and analyze patient data (Hashimoto, Rosman, Rus, & Meireles, 2018). This paper will discuss the application of AI to cervical, lung, and breast cancer treatment including the use of detection machines, segmentation techniques, and prediction algorithms, as well as weigh the challenges and social aspects of introducing AI to the medical field.

Figure 1. All types of Artificial Intelligence that will be discussed in this paper can be broken down as seen in the figure below. Within AI or all machines which are programmed to think in the way humans do, one can find machine learning. This is a more specific form of AI that is centered around the ability to improve performance based on past errors. A subcategory of machine learning is neural networks. Most studies that are discussed in this paper are on platforms and algorithms that are created with neural networks, or systems of algorithms that can use data to recognize patterns and then for this purpose, make predictions using that data. The final two categories, which are both subsets of neural networks, are deep learning and convolutional neural networks. Deep learning further imitates the way human brains think and learn. Convolutional neural networks are simply layered neural networks.

The Application of AI in Cervical Cancer Treatment

When dealing with AI and cancer detection, one of the most prominent issues that comes up is the invasiveness of diagnosis as well as how many cases are missed. While it can be cured if found at an early stage, many women die every year because their cancer was not detected early enough and symptoms did not appear until the cancer was too far advanced to treat. One of the few cancers that will be discussed in this paper is cervical cancer. Cells in the cervix can either be squamous cells, which when infected cause squamous cell carcinoma, or glandular cells which causes adenocarcinoma (P & M, 2018, p. 1). Because cervical cancer is difficult to detect and hard to treat if it has progressed too far, automated machines that can detect cervical cancer could significantly improve the survival rate of women suffering from the disease.

A study performed in 2018 proposed an automatic detection assisted by artificial intelligence which could detect cervical cancer in patients. The first was a preprocessing step and it involved taking a cervical cancer image and enhancing the contrast of it for better visibility using Oriented Local Histogram Equalization. Certain features such as roundness, sides, and circularity were then extracted from the image and used to train the neural network. The features were extracted to discriminate between a healthy cervical image and a cancerous one. The neural network classifier would then identify the cervical image as either benign or malignant by comparing it to the features used for training. For the classification of the tumour, a feed-forward backpropagation neural network was used to make the classification reach the highest possible accuracy. This type of neural network is built using three layers. The input layer accepts the elements of the features that were extracted. Three “hidden” layers in between all with different functions are also used, each of which has a different number of neurons, or calculated inputs from the previous layer. An example of this would be the average of all of the results from the previously hidden layer. The output layer is the response of the neural network and classifies the image as either normal or cancerous (P & M, 2018, p. 1).

Another important part of screening is finding cancerous lesions in images. Segmentation is a difficult but necessary part of this. The most common test for cervical cancer screening is cervical cytology or the pap smear test which screens for malignant tumour cells in the cervix. A positive cytology test can show different types of abnormal epithelial cells such as atypical squamous cells or atypical glandular cells. Segmentation is the process of separating masses in an image and is the most important step of cytology as it identifies cells based on their structures and morphology. In the majority of cervical cancer cases, cell segmentation is followed by abnormality classification which is frequently performed by feature-based machine learning algorithms as well as deep-learning approaches. Feature-based classification is based on feature extraction. Common features include the size, shape, colour, and texture associated with malignant tumours. Once feature extraction is complete, multiple different algorithms can be used for classification. A radial basis function support vector machine was developed that could classify images blocked into six different categories including blocks with many white cells, blocks with normal epithelial cells, and blocks with suspicious epithelial cells. The researchers expressed that the blocks with suspicious cells had a considerably different texture and colour features which set them apart from the others (Conceição, Braga, Rosado, & Vasconcelos, 2019, p. 21). The support vector is different from the layered neural network because instead of passing through a series of layers, there is only one function that separates or classifies the inputs into multiple categories. This method of classification skips the segmentation step entirely and saves a lot of time. Artificial neural networks, an unsupervised classifier, meaning they do not require inputs of labelled data to be trained, are another type of classifier that can study cell images and determine their level of abnormality.

Deep learning classification in the form of a convolutional neural network is a classification that can be performed without segmentation. On the other hand, this type of network does require far more computational time and high numbers of labelled data sets, making them impractical in clinical settings (Conceição, Braga, Rosado, & Vasconcelos, 2019, p. 22). 

With the help of such techniques, the survival rate can be increased and the chance of complications occurring can be decreased. These two measurements can also be predicted by Artificial Intelligence to ensure proper treatment and patient comfort. In an experiment to test survival rate prediction, a data set was collected from a total of 102 patients, all with cervical cancer that had already undergone initial surgical treatment. The researchers identified 23 demographic variables including age, BMI, and hormonal status, and 13 tumour-related parameters including tumour size and a number of lymph nodes, to direct the experiment. The computational intelligence methods that were applied had not yet been used to predict patient survival for cervical cancer treated by radical hysterectomy. Six of these were classifiers: Probabilistic neural network (PNN), multilayer perceptron network (MLP), gene expression programming (GEP), Support Vector Machines (SVM), Radial Basis Function Neural Network (RBFNN), and the k-Means method. The prediction ability of these models was determined by measuring accuracy, sensitivity, and specificity. The best results in the prediction of 5-year overall survival in cervical cancer patients who had already undergone radical hysterectomy came from the PNN model (Obrzut, Kusy, Semczuk, Obrzut, & Kluska, 2017, p. 4). The PNN model similar to the feed-forward backpropagation neural network mentioned earlier is made up of an input layer, a pattern layer, a summation layer, and an output layer. The PNN model along with other AI methods can be applied to various medical classification jobs (Fig. 2). 

The prediction of complications occurring during or after surgery is also vital to determine a patient’s chance of survival. One study was performed on 107 individuals with cervical carcinoma who had undergone surgery, and a cervical biopsy was taken to determine an AI algorithm’s ability to diagnose cancer. Complications around the time of surgery were evaluated both during the operation and post-operation. The gene expression programming (GEP) algorithm which makes and advances computer programs, was used for this study. The GEP was compared with the multilayer perceptron (MLP), the radial basis function neural network (RBFNN), and the probabilistic neural network (PNN), all of which are feed-forward neural networks. Each of the tested models was ranked based on their specificity, accuracy, and sensitivity. The highest accuracy was found in the MLP neural network. Complications near and around the time of surgery occurred in 47 patients although most of these were minor complications that did not severely harm the patient or put their lives in danger. Other more serious complications were found in 7 of those patients and included pulmonary embolism, or a gastric ulcer rupture (Kusy, Obrzut, & Kluska, 2013, p. 4). This study goes to show that it is imperative to identify any risk factors of surgery and choose the appropriate course of treatment as soon as possible because if products to remove cancerous tissues are postponed, the patient’s chance of survival is likely to decrease.

Figure 2. This figure is an example of a neural network that was created using a series of layers. A specific neural network that is similar to the one portrayed above is the feed-forward backpropagation neural network that was used for a 2018 study on the detection of cervical cancer. Another PNN model was also discussed in which there were three hidden layers in addition to the input and output layers. Patient scans were the input for this algorithm, like most other detection algorithms. The hidden layers included a pattern layer and a summation layer both of which had distinct functions and helped perform the calculations and an output, the prediction of 5-year patient survival, was reached.

The Application of AI in Lung Cancer Treatment

Like cervical cancer, lung cancer is life-threatening and is actually one of the leading causes of deaths in the world, therefore accurate diagnosis and treatment planning are extremely important for a patient’s survival. Recent breakthroughs of artificial intelligence, and specifically deep learning algorithms that can solve complex problems by analyzing images, are giving scientists hope. Researchers of one study developed a deep learning model to aid in lung cancer diagnosis to help reduce the work of pathologists. A convolutional neural network was trained to classify small patches of a histopathological image of a lung as either malignant or benign and had an accuracy rate of close to 90% (Wang et al., 2019, p. 8). This method would enhance the diagnosis of lung cancer by allowing for incredibly fast tumour detection when the region being studied is relatively small. Aside from diagnosis, the prognosis is one of the key parts of cancer treatment. Predicting if a tumour will recur and how long a patient will survive are crucial to determining the proper course of treatment. Wang’s team developed yet another CNN model that could segment slide images by the boundaries of the nucleus. Different features of the nucleus were extracted and used in a model that predicted the chance of recurrence (Wang et al., 2019, p. 9). 

Furthermore, scientists have found a relationship between a patient’s genetic files and pathological phenotypes, and genetic mutations that cause tumours. Such biomarkers are evolving and can be a useful tool in helping physicians with the screening and detection of lung carcinoma. An ideal biomarker is one that indicates biological, pathogenic, and pharmacologic processes and responses, and can impact clinical decisions in order to benefit a patient. Additionally, when being used for undefined pulmonary nodules, a biomarker should have the ability to predict and anticipate the diagnosis of cancer so that treatment can be administered as soon as possible and overdiagnosis is avoided. Scientists have established a few promising biomarkers (Fig.3), such as urine and saliva, that are currently used. Blood is another biomarker that can be used for lung cancer screening as it can help to identify and study the tumour and the space surrounding it, any metastases, and the patient’s immune response. Sputum, which comes from the airway epithelium can also be used for lung cancer and is able to supply data about any changes in a molecular structure close to the tumour cells. Autoantibodies are another form of biomarker which develop as a result of the formation of a tumour before any symptoms appear on images (Seijo et al., 2019, p.5). These autoantibodies have been discovered in all types of lung cancer, meaning in the future they could be indicators of lung cancer. Further studies are examining the rise of newer biomarkers that can be used alongside AI to decrease lung cancer patient mortality rates. A specific nano-array sensor which runs off of artificial intelligence, and has the potential to distinguish benign tumours from malignant ones, was used in a study to diagnose 17 different diseases from exhaled breath samples and there was an accuracy rate of over 85% (Seijo et al., 2019, p. 8). Other prediction models that use AI were also able to identify malignant tumours from harmless nodules, promising a bright future for AI-based diagnosis

AI platforms that use deep learning are being considered as a tool in fighting lung cancers is deep learning. Deep learning models allow researchers to remove certain characteristics for data that is imputed as well as have many layers and kernels, neurons in the layers between input and output layers, that allow them to perform many functions using the removed characteristics (Wang et al., 2019, p. 5). Deep learning has the ability to recognize complex data patterns, requiring no human input, and systems that use deep learning are not subjective the way human physicians are. A more specific class of deep learning is convolutional neural networks or CNNs. These models learn features from images and can eventually even predict outcomes. CNNs have been used in classification, segmentation, and detection, learning from histopathological and radiographic imaging, showing great potential in both areas. The automated feature extraction that deep learning models can complete is a huge advantage as manually removing features from pathology images is very time-consuming when the problem is challenging and complex or when researchers do not know very much about the input data and their relations with the outcomes that the model will predict. 

Like any disease, it is also always helpful if physicians are able to predict the chance of survival of a patient. A recent study was completed on medical images and information about tumours that could be helpful in prognostication efforts. 1194 individuals with NSCLC, who had been treated with either radiation or surgery, had CT scans taken of them and different elements that would determine a prognosis, Kwon as prognostic signatures, were detected using a convolutional neural network (Hosny et al., 2018, p. 1). CNN was highly successful in separating patients based on their chance of mortality. The network was also trained to predict the likelihood of a patient’s survival, 2 years after the start of their treatment. After the experiment was complete, the scientists dove in even deeper to get a better understanding of the different features detected by CNN and found specific areas that had the greatest impact on predictions made by the platform. To understand which regions in the CT images are responsible for the predictions made by the neural network, activation maps were created over the final convolutional layer. The intensity of the gradients in this layer determines how important each node actually was for the prediction. Most of what contributed to the prediction were from large areas of space both within and around the tumour, with higher CT density, and the areas with a lower density, such as the uncommon vessels, did not contribute very much to the predictions (Hosny et al., 2018, p. 12). Normal tissues such as bone tissue, which is of higher density, was ignored by the network as it appeared in most if not all images and had no significance on the tests. All of the actions that such a network takes- extraction, selection, prediction- are automated and have no data to back up why a certain prediction was made which makes it hard to prepare for failure. Although limitations do exist, there are possibilities for tools that can be created. An imaging tool with the ability to classify more specific information and identify treatment pathways would be helpful in managing all patients who suffer from NSCLC.

Lung cancer screening in developed countries is generally carried out with the use of LDCT or low-dose computed tomography. Although LDCT might be the favorable pathway for lung cancer screening and detection in the United States and other developed countries, developing countries face other challenges that makes it harder to integrate technology such as LDCT into routine clinical practice. It is very hard to develop programs that can screen for lung cancer in underdeveloped countries due to the vast amounts of pulmonary tuberculosis and chest infection cases. These conditions have similar symptoms to lung cancer such as fever, anorexia, weight loss, and cough, however individuals with histories of smoking, and a hoarse voice tend to be diagnosed with lung cancer (Shankar et al., 2019, p. 7). One of the most harmful consequences of using LDCT is that benign intrapulmonary lymph nodes and non-calcified granulomas are often hard to distinguish from pulmonary nodules, leading to many false positives, and thus unnecessary radiation which will eventually lead to the actual formation of cancer (Shankar et al., 2019, p. 7). Solutions for this issue include using computer aided diagnosis systems that are more sensitive with their detection, along with biomarkers that were previously mentioned, which can make screening more efficient. 

While LDCT is optimal, risks such as radiation and overdiagnosis along with cost make it hard for scientists to introduce at higher levels. Such methods that use AI have hopeful implementations in imaging and radiology, such as cancer detection and assistance making decisions, and the application of AI to pulmonary oncology will open up many pathways for diagnosis and prognosis using clinical, pathological, and morphological features of patient scans.

Figure 3. One of the things that can be used to determine if a patient has any genetic mutations are biomarkers. Examples of promising biomarkers are blood, sputum (mucus found in the airway epithelium), and autoantibodies. These biomarkers can be used to detect tumor cells because scientists are aware of how these biomarkers look in a normal state, however when a tumor has formed, the biomarker’s appearance will change due to surrounding cancerous cells.

The Application of AI in Brain Cancer Treatment

Another type of aggressive cancer that is hard to diagnose is brain cancer. When applied to data from MRI scans, AI has great potential in the field of neuro-oncology and is multi-purpose as it can help establish how harmful a tumor is, find invading gliomas, predict the chance of recurrence and survival of patients, assess the physician’s skills, and simulate cranial surgeries to strengthen neurosurgical training. 

Like in cervical cancer, segmentation is a large part of diagnosis of brain tumors through radiomics, and can be performed by deep learning AI platforms (Rudie, Rauschecker, Bryan, Davatzikos, & Mohan, 2019, p. 3). From these images, different features are extracted including size, shape, textures, and patterns. Machine learning platforms are then used to find relationships between the features, and determine the prognosis of the tumor. MR spectroscopy which compares the chemical composition of normal brain tissue with abnormal tumor tissue is used to classify gliomas and glioblastomas into different grades, depending on severity, and has the ability to identify regions where lymphocytic cells have invaded the tumor. Once found, the current treatment method for glioblastoma and gliomas is a resection of the tumor along radiation or chemotherapy using a medication called temozolomide (Rudie, Rauschecker, Bryan, Davatzikos, & Mohan, 2019, p. 8). However an effect that radiation and chemotherapy sometimes have is pseudoprogression, an increase in the size of the primary tumor or the appearance of a new lesion.

Machine learning devices that take patient images into account can be used in such instances to predict if a pseudoprogression is likely to occur. Researchers recently performed a study where an AI system was developed to outline characteristics of cancer cells in tissue grafts from patients that came from both the primary tumor and brain metastases, tumors in the brain that have formed from the original tumor. A live cell imaging algorithm was combined with AI and was used to study the movement of cells toward the area with damaged tissue and make out any differences between cells with and without brain metastatic potential. The study presented a device that could make calculations and predictions with the help of AI. The platform would be able to use a 3D measurement of cancer cells behavior in a BBB, or brain-blood barrier model outside of the organism and determine which cells have a brain metastatic characteristics. The visual differences between cancer cells that can form metastasis in the brain and those that cannot are very slight but the studies that used AI to identify these distinctions showed a large difference in the behavior of cancer cells and normal cells when they came across the BBB, making the AI device a very helpful tool for recognizing preseudopregrossions and potentially predicting them (Oliver et al., 2019, p. 4).  However there are always limitations that come along with medical breakthroughs such as these. While the ex vivo model in this experiment is able to identify differences in cells that are able to and not able to cross the barrier, the characteristics of cells with metastatic potential are still inconclusive. There are not yet enough features of a cell which platforms can detect that will allow an AI algorithm to accurately predict if a cell will metastasize. Furthermore, a brain cancer patient’s brain will have already changed in some way before diagnosis, making it more prone to the formation of metastases (Oliver et al., 2019, p. 8).

After it is properly developed, the use of AI to detect cells with the potential to metastasize in the brain will increase survival rate. Artificial intelligence can also be used to predict these chances of survival for patients suffering from cancer. A recent study used an artificial intelligence tool called the DeepSurvNet that runs off of neural networks in order to determine brain cancer patients survival rates and sort them into four different classes, based on just their histopathological images. To train the model, researchers used a dataset created using the medical records of brain cancer patients with 4 different types of brain cancer. 4 classes were used to classify patients by the time between their brain cancer diagnosis and death. Multiple regions of interest in the tumors from the imaging slides were also allocated to each of these classes. The model was then tested using completely new sets of data from histopathological slides. Glioblastoma tissue sections stained with H&E dyes of 9 new patients were analyzed. The device classified most patient samples in a single class, which was anticipated as the regions of interest are all taken from the same tissue sample (Zadeh Shirazi et al., 2020, p. 9). With the use of the DeepSurvNet classifier, physicians can use the difference between tumors that allow for different lengths of brain cancer patient survival to create specialized treatments and significantly decrease patient mortality.

In addition to it being hard to diagnose, because neurological cancer is such a rare condition, doctors do not get to see many patients with it and therefore lack training. Artificial Intelligence that can deliver feedback based on users touch, has the potential to create a realistic environment for trainees to practice their surgical skills without having to operate on real patients. Particularly, surgical simulations can be used in training for neurosurgery as the tasks required in the field are very technical and need to be performed under large amounts of pressure since one mistake could lead to severe consequences. A study on The Virtual Operative Assistant, shows the benefits of using AI to conduct training that tests cognitive skills and determines the level of psychomotor expertise that an operator possesses through the use of a surgical simulation. During the experiment, 50 participants, all with differing levels of expertise were recruited and classified into two groups: skilled and novice. The classification was completed after all of the participants were asked to complete a complex virtual reality simulation where they had to remove a brain tumor located beneath the pia mater subpial tissue, a type of connective tissue, using two devices-one in each hand. In order for the Virtual Operative Assistant to perform the classification, over 250 performance metrics were generated that were representative of differing levels of expertise in the surgical field and only 4 metrics with the highest level of accuracy were chosen after careful consideration and selection processes (Mirchi et al., 2020, p. 4). After the machine learning algorithm computed its classification between “skilled” or “novice”, it also gave users a breakdown of its assessment on both the safety and movement metrics, and rather than assessing each metric on its own, the Virtual Operative Assistance included the relationship between metrics, allowing students to recognize that one strong metric may be making up for poor performance in another one. The three forms of feedback that the Virtual Operative Assistant is able to give users, auditory, text, and video-based, is what makes it extremely beneficial in the world of science. This new technology enables scientists to understand the expertise of an individual, identify cognitive expertise in tasks that are much too complex for human teachers to notice, and mimic real life training, all making it the perfect tool for simulation-based learning.

Evolving technologies and Further Use for Medical Education

Simulation-based training systems, such as the Virtual Operative Assistant, are able to develop checklists that evaluate different skills using machine learning algorithms (Sapci & Sapci, 2020). While there are numerous applications for AI cancer treatment: screening and detection, survival rate prediction, and surgical simulations which allow doctors to more efficiently develop surgical skills and treat patients, AI platforms do possess many challenges that come along with using them. Of the different challenges that AI platforms in medical training pose, feedback and liability issues are two of the most prominent. A study done at Mount Sinai Hospital created a type of AI technique known as deep learning using data from 700,000 patients (Paranjape, Schinkel, Nannan Panday, Car, & Nanayakkara, 2019). Their algorithm was highly accurate and was able to diagnose conditions that even experts struggle to diagnose, such as schizophrenia. However, AI systems often lack the ability to provide users with a proper explanation for how a certain answer or prediction was reached (Fig. 4a). These algorithms cannot properly understand the cognitive thinking of learners and therefore cannot properly train them in the actual areas where they need to be trained. This brings about the issue of liability because it becomes very hard for patients to trust a system that cannot provide an explanation for how it reached a diagnosis, and if a calculation were to be made incorrectly that puts a patient in danger, then it is not known whether the doctor, the hospital, or the company that developed the AI device is liable (Paranjape, Schinkel, Nannan Panday, Car, & Nanayakkara, 2019).

Because of their ineptitude to comprehend the emotional reasoning of users and provide appropriate feedback, AI-powered teaching platforms enable students to “cheat” the system. Many algorithms in artificial intelligence teaching tools do not actually train surgeons and increase their skill level. Rather, they make the assumption that a student is skilled in a certain area because they were able to accomplish one certain task. In the case of the Virtual Operative Assistant, this ability to cheat (Fig. 4b) can be credited to the relatively broad parameters that classify students as either skilled or novice (Mirchi et al., 2020, p. 16). In the experiment completed using this specific AI platform, there was a misclassification where 4 participants that were actually at the novice level were labeled as skilled. Such errors make it difficult to trust AI-powered teaching tools and implement them into routine medical practice and surgical training.This is where human expertise comes in and proves essential in the learning process. If AI platforms undergo diligent training alongside human experts that can properly assess the algorithm and recognize specific markers of a good surgeon, then cheating the system would be much less likely. Furthermore, the issue of learners feeling a disconnect from their teacher due to lack of feedback and properly backed up explanations, which can actually damage a students skill level, can be resolved by human interaction (Chan & Zary, 2019). AI could be substantially more useful for tasks such as computerised testing and cancer screening or diagnosis, but if physicians and AI-based machines are able to work in harmony then patient outcomes are guaranteed to improve as AI has the potential to process large amounts of data including medical reports, notes from pharmacists, and genetic reports, as well as analyze all of it. Nonetheless, a major thing that it cannot take the place of is the beauty of doctor-patient and doctor-student interactions (Paranjape, Schinkel, Nannan Panday, Car, & Nanayakkara, 2019).

Figure 4a. This figure demonstrates the lack of explainability in “blackbox” algorithms. The term “blackbox” means that while the algorithm is producing an educated response, scientists are not able to access the process being completed inside the algorithm. The algorithm does not give them feedback so doctors are not able to provide patients with adequate information to back up the reasoning behind their decisions. In this figure, the same patients scans are being input into two algorithms and both come up with the same patient diagnosis however one algorithm allows doctors to have an understanding of the decision making, while the other is a complete “blackbox”. Due to their multilayer non linear structure the Virtual Operative Assistant, a deep learning model, is considered a “blackbox” and criticized for its lack of transparency.
Figure 4b. A potential pitfall that can occur with these “blackbox” algorithms is the possibility of cheating. In the case of the Virtual Operative Assistant as shown below, the parameters for classification as either skilled or novice are relatively broad. For example, if the algorithm determined that a participant’s performance was ‘poor’, they would most likely be classified as novice. However, due to the broad parameters, there is still likelihood that they could be classified as skilled. In situations with physicians and patients where the doctor has to make a life-saving decision, the possibility for such a mistake is unacceptable and can prove to be detrimental.

Doctor-Patient Feedback and Interpretation

AI in the healthcare field is expected to grow rapidly in the years to come, but with growth comes limitations, which is why it is crucial for AI to be implemented into the healthcare system with ethical and legal aspects in consideration. A large imitation is that AI systems do not have feelings and can’t care for or have sympathy for patients in the same way that doctors do. The “quadruple aim” of healthcare consists of improving experience of care, improving health of populations, reducing per capita costs of healthcare, and improving the work life of healthcare workers (Kelly, Karthikesalingam, Suleyman, Corrado, & King, 2019, pg. 1). But healthcare systems are struggling to meet these goals.

The FDA has already cleared close to 40 devices that run off of AI and can be used for medical purposes. one of these is the IDx-DR, a system that can output a screening decision without the help of human interpretation of the image or results. The system then recommends the physician either rescreen or refer them to a specialist (Gerke, Minssen, Cohen, Bohr, & Memarzadeh, 2020, pg. 5). However while AI can improve imaging, diagnosis, and surgery, it will be difficult to manage AI when informed consent is considered. It is a common question whether it is the physician’s responsibility to inform the patient about AI and the way it works, and if the doctors have to even inform the patient at all that AI is being used. Some argue that it doesn’t matter how an AI system reaches its prediction, but more important is if the decision is correct but this can cause an issue in certain cases as many machine learning algorithms are known as “black boxes”, even the inventor does not know how the program reaches its final decision. The datasets being used to train the algorithms also need to be reliable, trustworthy, valid, and effective- the better the training data, the better the accuracy of the AI algorithm. Even after the first model is developed, the program will need further tweaks to be made. This includes data bias. Many AI algorithms have proven that they do have a bias when dealing with ethic origin, gender, age, or disabilities. These biases could lead to false diagnosis and jeopardize the safety of patients by making treatments ineffective (Gerke, Minssen, Cohen, Bohr, & Memarzadeh, 2020, pg.10). If an AI algorithm outputs a recommendation for treatment that a human physician would not have picked, therefore making it wrong, and the physician decides to use it anyways and it harms the patients, then it is likely that the physician is at fault for medical malpractice. It is also important to think about hospitals and if it becomes their fault that they bought and applied the AI systems to their practice, but this is why AI should be used more for assistance with decision making, rather than fully depended on.

For patients suffering from kidney failure, or end-stage renal disease, renal transplants are the best option for a patient to survive, yet dialysis is often a more reasonable choice due to the shortage of organ donors. However, currently dialysis software is not equipped to respond to unanticipated events that can occur during dialysis treatment. Miniature artificial kidneys that can provide a personalized dialysis treatment and are capable of real-time monitoring are currently in the process of being developed (Hueso et al., 2018, pg. 5). Data analytics supplied by fields of artificial intelligence such as machine learning and computational intelligence are expected to principally play a role making sure these new dialysis technologies are both efficient and safe for patients. Due to the complexity of the technologies involved in the creation of these dialysis machines, there are challenges for implementation of the devices in healthcare and biomedicine. Data analytics along with AI provide the baseline for medical decision support systems, but application of these AI algorithms in the medical field has its challenges, the biggest being the ethical issue (Hueso et al., 2018, pg. 2). Although this device will make the jobs of medical professions easier, it may make interactions with patients uncomfortable and lose trust in their doctors. For example, these automated devices do not have the ability to explain the reasoning for their decision and empathize with patients the same way that doctors can, making it hard for patients to understand their own course of treatment. AI algorithms are not able to express the relationship between the data they have observed and the outcomes that have been formed as a result of it. 

While it is possible for an algorithm to overcome all of these challenges, human-computer interactions are one of the key aspects in gaining a better understanding of the way algorithms interpret data. Multiple algorithms produce great results but lack the ability to explain why they landed at those particular results. Even if scientists are able to understand the math that is involved in creating the algorithm, it is virtually impossible to determine which model made a specific decision. This is problematic as it has rendered many algorithms as untrustworthy, uninterpretable and unexplainable. Therefore there is a tradeoff between performance and expandability. Deep learning models have a very high level of performance but they are hard to interpret while linear regression models decision trees are relatively to interpret but have poorer decision making skills. Interpretability of AI algorithms is the ability of a human to understand the way it made a connection between features extracted and its predictions. Approaches to solving this interpretability issue can be categorized into two categories: whether they need internal information such as parameters to operate (also known as the level of transparency) or the amount of accessibility there is to the internal information of the model. Methods that require access to the internal information are considered to be working on “white boxes” (Reyes et al., 2020, pg. 2). An example of this is a CNN or convicted neural network where a radiologist uses a given layer of network to create a map which can be laid on an image that shows which regions are important for predicting if the patient has a disease. Black boxes on the other hand do not need access to this internal information and instead work directly with the input and output of the model to analyse how changes to the input will change the output. There are multiple visual techniques that give insights into the way that AI algorithms behave and they arrive at certain decisions. Two of these basic approaches are partial dependence plots and individual conditional expectation plots. Both methods are used to interpret black-box models and show the way a model’s predictions are dependent on the features. This helps predict which features will change the prediction when their value is changed (Reyes et al., 2020, pg. 2).

The goal of interpretability isn’t exactly to understand exactly how an AI system works, but to have enough information to understand it to the best extent possible and it is not always necessary. A wrong diagnosis in radiology can lead to extreme consequences for a patient, but reading images is prone to interpretation errors. Interpretability is a fast evolving field that has been at the center of AI research with great potential for future development of safe AI technologies (Kelly, Karthikesalingam, Suleyman, Corrado, & King, 2019, pg. 1). But before AI can be implemented into various tasks within radiology, task-specific interpretability solutions are required, and if algorithms known as “black boxes” are used in medicine, they need to be used with a great deal of judgement and responsibility. AI developers should be aware of the many consequences of algorithms and unintentionally lead to and make sure they are created with all patients in consideration. Doctors and surgeons being involved in this process can increase its efficiency significantly. If the interpretability of algorithms can be improved, then human-algorithm interaction would be smoother and the future adoption of AI with consideration of data protection, fairness and transparency of algorithms, and safety, would be supported by a large number of physicians.

The Impact of AI on Jobs

While many physicians may support the implementation of AI, no machine can work at its full potential without the presence of doctors, but studies have shown that medical students are not spending enough time getting acquainted with newer technologies that involve Artificial Intelligence. Currently, medical education is centered around 6 major areas: medical knowledge, communication skills, patient care, professionalism, practice based learning, and systems- based practice (Paranjape, Schinkel, Nannan Panday, Car, & Nanayakkara, 2019). Most of this training focuses on taking in large amounts of information and applying this information to patient care- a process based mostly on memorization. In order to improve outcomes in clinical settings, students need to learn how AI functions and how it can augment their work. The many promises of AI include automated image segmentation, detection of cancerous lesions, and comparison of images. While it can be fatiguing for human pathologists to detect small traces of cancer on a slide, AI systems are not affected by this and can scan a number of slides without losing their accuracy. AI can also help physicians to improve the quality of patient care by taking care of repetitive tasks and tedious tasks and managing large amounts of data, in addition to being another opinion for making decisions.

With AI algorithms showing such great amounts of promise in radiology, pathology, and cardiology, a question that arises is while AI algorithms replace human physicians? Recent data expresses that when considering its image and predictive analysis, AI might soon prove to be more efficient than radiologists. However, it is likely that AI will not replace the role of general physicians, but rather augment them. For example, an AI system is able to take over the job of a factory worker who performs a certain task repeatedly, but in the case of replacing medical professionals, AI is unable to engage in interactions with patients that are crucial in gaining their trust, and restoring them. One study on this topic deals with breast cancer. It suggests that digital mammography is not perfect and only has a sensitivity of around 85%. The other 15% that is not detected is a result of human error- what radiologists are able to identify on scans. Furthermore, the question of whether or not this practice is ethical is an important one. While replacing human workers with AI systems, it may benefit the economy as a whole, but the effect that it has on individuals whose jobs have now been taken away, is detrimental.

In most cases, technology is designed to perform a specific task which changes the demand that workplaces have for certain skills. These changes can influence the skill requirements, social well-being of workers, and career mobility, for different occupations. Limitations on data to train AI algorithms will restrict these skill pathways, but scientists can surmount this obstacle by prioritizing data collection that is detailed and responsive to real-time changes in the labor market (Frank et al., 2019). This improved data collection will enable the use of new tools that rely on data, including machine learning systems that more accurately reflect the complexity of labor systems. New data will also lead to new research that will strengthen our understanding of the impact of technology on the supply of and demand for labor

However, AI systems do tend to result in a number of false positives, resulting in extreme measures being taken without certainty of harmful cancer presence in the body. This is where radiologists are still crucial in the medical field, even with the presence of AI. Furthermore, false positives are an issue which can lead to anxiety, along with unnecessary biopsies and tissue removal(Ahuja, 2019). AI has the potential to assist and augment physicians rather than replacing them entirely by combining data and providing help in the decision making process by recommending certain treatment options. AI can also remove some of the burden of work from physicians by performing tedious tasks. Speech recognition is another useful device that can replace keyboards and decision management can help physicians to make more informed decisions that take into account both patient outcome and cost of treatment (Paranjape, Schinkel, Nannan Panday, Car, & Nanayakkara, 2019).

Figure 5. As AI continues to develop, the impact it has on jobs will grow increasingly. This is because AI algorithms have a set of skills which humans do not. For example, AI algorithms will never get tired, they are able to perform tedious tasks without the quality of the job going down. AI algorithms are also able to detect tumors in a way the human eye cannot and can perform calculations much faster than the human brain. All of these skills allow algorithms to make informed decisions and predictions. Because of their ability to make informed decisions, AI has begun to take the jobs of receptionists and programmers who do tedious work. AI also has the potential to take the jobs of physicians however the lack of doctor-patient interactions may make patients uncomfortable in a hospital setting. Taking jobs away from human laborers will also have a devastating impact on the US economy in the long run.

Ethical Considerations

Due to the fact that the field of AI is a relatively new form of technology, it’s implementation in the real world raises a number of questions about the ethical side of the technology. Within the AI algorithms themselves, one prominent issue is model bias. The data that is being used to train such algorithms has the potential to be influenced by multiple outside factors including a bias towards the humans who collected the dacollectors of the data. As a result, an algorithm may be biased towards a specific group when predicting whether or not an individual should receive a certain treatment. It is important for researchers to consider this aspect of AI and work towards mitigating the effect of such biases. Data not representative of a large population can result in a model that is biased to subjects highly prevalent in the data set. In addition, for the highest level of fair and accurate model performance, it is imperative for scientists to split data so that platforms can be tested with images separate from the training data.

The first issue that seems to directly influence patients is how although AI may not entirely replace doctors, it will significantly alter relationships between patients and their physicians and nurses. Many companies that distribute electronic health records however have overlooked this disadvantage and have focused on only the positive aspects including how AI will be able to simplify interactions with complex data and reduce the time it takes to complete tasks. However to many patients, it is incredibly important for their own comfort and satisfaction to maintain relationships with their doctors. If AI algorithms are set to take over scheduling appointments, making payments, and even running follow-up visits, then this doctor-patient interaction time will be compromised. Furthermore, it is important to take into account the immense amount of data that algorithms require access to for training. While the majority of companies are sure to keep their data protected in order to abide by HIPAA, a privacy law that creates national standards to protect personal medical data, some organizations do allow their data to move freely in and out of their company. This sacrifices patient privacy and security in a way that didn’t have much of an effect before AI was implemented into medicine. Finally, the legal responsibilities that come along with having a hospital run by AI are vast. For any negative consequences that could have been fixed previously, oftentimes at first glance it seems that it is the responsibility of the provider of the AI algorithm. Providers do need to be certain that their algorithms they are providing to hospitals use relevant and accurate data that can make decisions in the most beneficial way possible, but questions surrounding this topic remain unanswered. One could also argue that negative outcomes are the doctors’ fault because they relied too heavily on an algorithm instead of using their own expertise to make a decision. In the end, it is the responsibility of contributors- providers of AI, developers of AI, patients, doctors, and all others involved in the process- to make sure artificial intelligence develops in the medical field in a safe and ethical manner.

Conclusion

Through its multitude of uses across the field of medicine and oncology specifically, AI has the potential to transform the way physicians work, and the way patients are treated. Within cervical cancer, lung cancer, and breast cancer, AI algorithms are able to detect lesions in scans and use segmentation techniques to separate masses within such an image and identify cells based on their structures. Furthermore, they have the ability to utilize these images and extracted features in order to classify images and predict the possibility of recurrence and even chance of patient survival. There are currently a number of platforms being developed which can perform these tasks and a number more being developed to teach medical students how to interact with technology and practice their skills in a real life setting before actually doing so. In the future, AI is set to enable faster and more accurate diagnosis, reduce human errors that are a result of fatigue, complete repetitive and tedious labor taste, decrease medical costs, perform minimally invasive surgery, and increase survival rates. Specific examples of  prospective applications for AI as the field grows, are in analyzing relationships between patient outcomes and treatment administered to patients, diagnosis, forming protocol for certain treatments, personalized medicine, and patient care. However, despite these fascinating advancements in technology and medicine and the tremendous potential AI has to revolutionize medicine, there are still some things that AI is not able to accomplish. It lacks the ability to have social interactions with patients in a way that humans doctors can, and will continue to take jobs from employees around the world as its role expands. In order to surmount these obstacles, scientists will have to consider how far patients are willing to go in regards to putting trust in their doctor as well as the economic impact of AI and how it could in turn harm the economy by taking away large numbers of jobs. In order to improve success of AI in the fields of cancer detection, diagnosis, and treatment, these factors must be taken into account.

Sia Shah, Youth Medical Journal 2021

References

Ahuja, A. (2019, October 4). The impact of artificial intelligence in medicine on the future role of the physician. Retrieved November 18, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6779111/

Bi, W., Hosny, A., Schabath, M., Giger, M., Birkbak, N., Mehrtash, A., . . . Aerts, H. (2019, March). Artificial intelligence in cancer imaging: Clinical challenges and applications. Retrieved November 18, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6403009/

Chan, K., & Zary, N. (2019, June 15). Applications and Challenges of Implementing Artificial Intelligence in Medical Education: Integrative Review. Retrieved August 10, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6598417/

Conceição, T., Braga, C., Rosado, L., & Vasconcelos, M. (2019, October 15). A Review of Computational Methods for Cervical Cells Segmentation and Abnormality Classification. Retrieved October 02, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6834130/

Frank, M., Autor, D., Bessen, J., Brynjolfsson, E., Cebrian, M., Deming, D., . . . Rahwan, I. (2019, April 2). Toward understanding the impact of artificial intelligence on labor. Retrieved November 03, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6452673/

Gerke, S., Minssen, T., & Cohen, G. (2020). Ethical and legal challenges of artificial intelligence-driven healthcare (1010864836 778175037 A. Bohr & 1010864837 778175037 K. Memarzadeh, Eds.). Retrieved October 27, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7332220/

Hashimoto, D., Rosman, G., Rus, D., & Meireles, O. (2018, July). Artificial Intelligence in Surgery: Promises and Perils. Retrieved November 18, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5995666/

Hosny, A., Parmar, C., Coroller, T., Grossmann, P., Zeleznik, R., Kumar, A., . . . Aerts, H. (2018, November 30). Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study. Retrieved September 15, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6269088/

Hueso, M., Vellido, A., Montero, N., Barbieri, C., Ramos, R., Angoso, M., . . . Jonsson, A. (2018, February). Artificial Intelligence for the Artificial Kidney: Pointers to the Future of a Personalized Hemodialysis Therapy. Retrieved October 27, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5848485/

Kelly, C., Karthikesalingam, A., Suleyman, M., Corrado, G., & King, D. (2019, October 29). Key challenges for delivering clinical impact with artificial intelligence. Retrieved October 27, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6821018/

Kusy, M., Obrzut, B., & Kluska, J. (2013, December). Application of gene expression programming and neural networks to predict adverse events of radical hysterectomy in cervical cancer patients. Retrieved October 02, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3825140/

Mirchi, N., Bissonnette, V., Yilmaz, R., Ledwos, N., Winkler-Schwartz, A., & Del Maestro, R. (2020, February 27). The Virtual Operative Assistant: An explainable artificial intelligence tool for simulation-based training in surgery and medicine. Retrieved August 10, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7046231/ 

Obrzut, B., Kusy, M., Semczuk, A., Obrzut, M., & Kluska, J. (2017, December 12). Prediction of 5-year overall survival in cervical cancer patients treated with radical hysterectomy using computational intelligence methods. Retrieved October 02, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5727988/

Oliver, C., Altemus, M., Westerhof, T., Cheriyan, H., Cheng, X., Dziubinski, M., . . . Merajver, S. (2019, March 27). A platform for artificial intelligence based identification of the extravasation potential of cancer cells into the brain metastatic niche. Retrieved August 24, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6510031/ 

Paranjape, K., Schinkel, M., Nannan Panday, R., Car, J., & Nanayakkara, P. (2019, December 3). Introducing Artificial Intelligence Training in Medical Education. Retrieved August 13, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6918207/

P, E., & M, S. (2018, December 25). Automatic Approach for Cervical Cancer Detection and Segmentation Using Neural Network Classifier. Retrieved October 02, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6428557/

Reyes, M., Meier, R., Pereira, S., Silva, C., Dahlweid, F., Von Tengg-Kobligk, H., . . . Wiest, R. (2020, May 27). On the Interpretability of Artificial Intelligence in Radiology: Challenges and Opportunities. Retrieved October 27, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7259808/

Rudie, J., Rauschecker, A., Bryan, R., Davatzikos, C., & Mohan, S. (2019, March). Emerging Applications of Artificial Intelligence in Neuro-Oncology. Retrieved August 24, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6389268/

Sapci, A., & Sapci, H. (2020, June 30). Artificial Intelligence Education and Tools for Medical and Health Informatics Students: Systematic Review. Retrieved August 13, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7367541/

Seijo, L., Peled, N., Ajona, D., Boeri, M., Field, J., Sozzi, G., . . . Montuenga, L. (2019, March). Biomarkers in Lung Cancer Screening: Achievements, Promises, and Challenges. Retrieved September 15, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6494979/

Shankar, A., Saini, D., Dubey, A., Roy, S., Bharati, S., Singh, N., . . . Rath, G. (2019, May). Feasibility of lung cancer screening in developing countries: Challenges, opportunities and way forward. Retrieved September 15, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6546626/

Wang, S., Yang, D., Rong, R., Zhan, X., Fujimoto, J., Liu, H., . . . Xiao, G. (2019, October 28). Artificial Intelligence in Lung Cancer Pathology Image Analysis. Retrieved September 15, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6895901/
Zadeh Shirazi, A., Fornaciari, E., Bagherian, N., Ebert, L., Koszyca, B., & Gomez, G. (2020, May). DeepSurvNet: Deep survival convolutional network for brain cancer survival rate classification based on histopathological images. Retrieved August 24, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7188709/

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s