7 ways AI revolutionizes cancer prediction, detection, and treatment

AI and machine learning can help with cancer prediction. Artificial intelligence can detect existing cancers and identify people at high risk of developing the disease before it sets in. This allows doctors to monitor these patients closely and intervene immediately when needed. Regina Barzilay, a computer scientist at the Massachusetts Institute of Technology (MIT), wanted to test AI in predicting cancer. The MIT team studied its ability to flag women at risk of breast cancer before they develop any tangible signs. She gathered around 89,000 mammograms of approximately 40,000 women taken over four years and checked the scans against the national tumor registry to determine which patients got cancer. Afterwards, Regina trained a machine learning (ML) algorithm, a subtype of AI, on a subset of these images and ran it to make predictions. The algorithm accurately classified 30% of women who developed breast cancer in the future as a high-risk group. In comparison, human physicians using the standard Tyrer-Cuzick model flagged only 18% of the patients.

There are many applications of AI in medical imaging. Detecting and classifying cancerous tumors is among the most noticeable. In September 2021, the FDA approved an AI-powered pathology product for cancer, Paige Prostate. This AI tool helps detect prostate cancer by working together with FullFocus digital pathology viewer. As a precondition of this approval, the FDA studied data from a clinical experiment where 16 pathologists examined 527 prostate biopsy images searching for cancer signs. Each doctor performed two assessments, one with the help of Paige Prostate and one unassisted. This experiment demonstrated that this AI-based tool improved the cancer detection rate by 7.3% on average. Additionally, it reduced false-negative diagnoses by 70% and false positives by 24%. In another instance, Mark Schiffman, Senior Investigator at National Cancer Institute, worked with his colleagues to build an AI-based algorithm, which can look at cervical tumor images and identify precancerous changes that require immediate medical attention. The model was trained on 60,000 cervical images, and as Mark puts it, “The computer algorithm was at least twice as accurate as the best doctors we showed the images to.” By analyzing cervical abnormalities, the AI algorithm could predict which patients would develop cervical cancer seven years in the future.

AI-enhanced blood tests can help doctors detect cancer more accurately. An article published in Cancer Cell International claims that blood profiling, where AI algorithms analyze ctDNA and miRNA plasma profiles, is a superior method of detecting and monitoring cancer compared to regular CT scans. Researchers at the Johns Hopkins Kimmel Cancer Center developed a novel AI-based technology for diagnosing lung cancer through blood tests. This approach was tested on blood samples of 796 subjects in the US, Denmark, and the Netherlands. Researchers combined this blood test with protein biomarkers, patients’ clinical risk factors, and CT scans. As a result, they accurately spotted cancer in 91% of subjects with early disease stages and in 96% of patients with advanced cancer phases.

ML in cancer detection can help individuals get initial feedback on their abnormalities through self-diagnosing and without the need to make an appointment with the doctor. Diagnosis using such tools is not necessarily final. They are subject to pathologists’ approval. SkinVision, based in Amsterdam, developed a mobile app that assists users in screening their skin abnormalities for cancer. They can capture a photo of a suspicious spot on their skin with a smartphone camera and submit it for evaluation. An AI algorithm analyzes the color, texture, and the shape of the lesion and gives feedback to the user within 30 seconds. The algorithm is believed to offer 95% accuracy in skin cancer detection. Users who were deemed at high risk of cancer are referred to a pathologist to confirm the diagnosis. Another example of AI in cancer detection through self-diagnosing comes from Google. Their tool is called Dermatology Assist, and it can identify abnormalities on skin, nails, and hair. Researchers trained it on 65,000 images, and it can identify 288 skin conditions. In addition to submitting skin photos, this AI app asks users to fill in a questionnaire online.

AI application in immunotherapy mainly focuses on evaluating the effects of different treatments and helping doctors adjust their prescriptions. A research team at UT Southwestern Medical Center and MD Anderson Cancer Center built an AI-powered technique for identifying which neoantigens (peptides produced by mutations in cancer cells’ genomes) are recognized by a patient’s immune system. Such AI algorithms would allow predicting cancer cells' response to immunotherapies. Our immune system’s T cells are constantly watching for signs of cancer and other invading bodies. When these cells recognize neoantigens, they bind together. However, some neoantigens remain unrecognized, allowing cancer to grow. There are tens of thousands of types of neoantigens. Analyzing their ability to trigger T cells’ response is a tedious, costly, and time-consuming task. With the help of machine learning, this is becoming possible. Here is what Tao Wang, PhD, Assistant Professor, Population and Data Sciences at UT Southwestern Medical Center, said about this matter, “Determining which neoantigens bind to T cell receptors and which don't has seemed like an impossible feat. But with machine learning, we're making progress." Having this knowledge would allow researchers to develop personalized T cell-based therapies and cancer vaccines, and it will help predict patients’ response to immunotherapies.

Various types of cancer may have different reactions to the same drug. AI can predict how cancerous cells will behave when treated with different compounds. This knowledge helps develop new anticancer drugs and understand when to apply them. For example, a team of researchers developed a random forest algorithm, which can forecast anticancer drugs’ activity based on the cancer cell’s mutation state. In another example, researchers at Aalto University, the University of Helsinki, and the University of Turku in Finland built an ML-powered model, comboFM, that can determine which combination of drugs has the most potential to kill particular cancer cells. ComboFM uses historical data from experiments performed on similar cells and drugs when looking for the optimal combination. The researchers claim their software can serve as an effective way for pre-screening drug combinations for different oncology applications. The influence of AI in cancer treatment expands to generic drugs as well. Laura Kleiman, the founder of Reboot Rx, worked with her team to build an AI-powered solution that would determine which generic drugs have the potential for treating cancer. Currently, Laura focuses on prostate cancer. She employed Reboot Rx AI to analyze literature describing clinical studies of non-cancer drugs ever tested to treat prostate cancer. The algorithm itself can determine the relevance and significance of a study.

Enhancing genome sequencing with AI can help in tumor characterization and personalized treatment development. When it comes to lung cancer, pathologists often struggle to detect and distinguish between two common tumor types, adenocarcinoma and squamous cell carcinoma. Scientists at New York University’s Langone Medical Center trained a deep learning algorithm developed by Google on 1,600 histopathology slides of lung tumors available on the Cancer Genome Atlas. After testing the algorithm, the team said it was as accurate as an experienced pathologist. Still, it could deliver results in a matter of seconds, while a human doctor would need a few minutes. Recently, Mayo Clinic announced its collaboration with Personalis to offer comprehensive cancer genomic sequencing to patients and their healthcare providers. They can use this data to develop new personalized cancer treatments and make more informed decisions.

Generally, AI has many benefits in healthcare. Here are four main advantages of incorporating artificial intelligence in cancer treatment and detection:

AI, supported by big data, enables doctors to study diverse information about the patient and the cancer cells coming up with personalized treatments. This type of therapy will result in fewer undesirable side effects. It will have a higher impact on cancer cells and cause less damage to the healthy ones. Cedars Sinai Cancer teams up with Tempus, a Chicago-based AI and precision medicine provider, to create molecular twins of cancer patients using artificial intelligence in cancer treatment. These twins are virtual replicas of those humans. They include information such as DNA, RNA, and proteins and help identify the most effective approach to cancer treatment for a particular individual.

Using AI in cancer detection will improve the accuracy of diagnosis, reducing false positives and negatives. Research on breast cancer detection gives us proof. While examining mammograms, doctors deliver false-positive news to one in ten of their female patients, and these women have to deal with the stress and unnecessary invasive procedures. Google’s research team built an AI-powered software that cut false positives in mammogram readings down by 6% and false negatives by 9%. Another group of scientists developed an AI algorithm for breast cancer detection. During an evaluation, this model helped radiologists reduce false-positive rates by 37.3%.

AI helps radiologists identify which tumors/abnormalities are malignant and actually need treatment. A study published in the Journal of the National Cancer Institute describes how AI algorithms can detect precancerous lesions in cervical images and distinguish them from other abnormalities, thereby avoiding overtreating patients for minor issues.

Sometimes, only after the removal surgery is performed, the doctors realize the tumor was benign, and the surgery could have been avoided altogether. With the help of AI in cancer detection, such incidents can be reduced significantly. For instance, one study revealed that AI can reduce breast-conserving surgeries by 30.6%. Machine learning algorithms can be trained to identify cancerous lesions using image-guided needle biopsies. The researchers examined 335 prospective cancer patients with a random forest ML algorithm and observed that it prevented one-third of unnecessary surgeries. In another example involving brain tumors, researchers from Harvard University and the University of Pennsylvania developed a deep learning algorithm for tumor classification. It can identify and characterize the isocitrate dehydrogenase (IDH) mutations from MRI images of gliomas without the need for invasive procedures that would be required otherwise.

Unfortunately, bias is not uncommon among AI models. They can display discrimination against certain ethnic groups and even against hospitals. Algorithms that work well for one care center might drop in performance when transferred elsewhere. A research team from the University of Chicago demonstrated how a machine learning-powered cancer detection application taught itself to consider the medical institution submitting the image as a factor in determining whether the scan shows signs of cancer. The good news is that you don’t always need to adjust AI algorithms for different population groups. For example, researchers from MIT’s Computer Science and AI Laboratory and Massachusetts General Hospital developed an ML model for breast cancer prediction. According to Allison Kurian, an Associate Professor of Medicine and Health Research at Stanford University School of Medicine, this algorithm performs equally well for both white and black patients. To fight bias, developers can check the training datasets on preexisting biases. There are computational methods that can detect and mitigate bias in data. Still, it is best to validate AI applications against population segments that the training data didn’t represent. Also, it is worth conducting regular audits to make sure bias doesn’t sneak into ML algorithms as they continue to learn.

Some pathologists are afraid that by working with AI on cancer treatment, they are simply training their replacement. They hear quotes, such as, “An AI algorithm can learn from a much larger library than a radiologist can. In some cases, a million images or more.” And they start worrying and thinking that AI can surpass them in everything they do. The reality is that AI can be great at one task or at a few tasks, but it will not replicate pathologists’ scope of work. Researchers who interact closely with AI believe this technology is meant to complement doctors, not to replace them. The quote presented above is said by John Shepherd, Professor and Researcher at the University of Hawaii. Here is how he ends his idea: “Thus, it could solve the endemic shortage of highly trained radiologists, which is very severe here in Hawaii.” For Shepherd, AI is playing a supporting role, helping when radiologists are not around. Some researchers even believe that human pathologists are more accurate than AI in cancer detection. Scientists at the University of Warwick, UK, studied the performance of AI in breast cancer detection and therapy. They evaluated 12 existing studies and found that 94% of AI models presented in these studies were less accurate than a single radiologist. And all of them lost the battle against a team of two or more doctors.

Healthcare data is typically stored in heterogeneous and unstructured ways, and it is challenging to standardize terminology across medical facilities. Initiatives, such as patient-reported outcome measures (PROMs), allow collecting standardized data from the start when patients experience distress. However, putting pressure on physicians to collect more data can lead to burnout and other increased workload problems. PotentiaMetrics, a healthcare analytics company, employed ITRex to develop an AI-enabled web portal for collecting, managing, and presenting patient data. The platform was intended for cancer patients. It gathered and maintained their information from the moment of diagnosing and throughout their survivor journey. Patients would enter their data using a web-based questionnaire and generate useful reports. Also, the platform helped physicians craft personalized treatment plans based on the detailed information they received.

More training data leads to higher system accuracy, and this is a challenge in the healthcare sector. The existing medical image datasets are significantly smaller than natural image sets. For example, LUNA dataset of CT images contains only 888 instances, and the Indian Diabetic Retinopathy Image Dataset (IDRiD) includes approximately 600 images. Working with a large properly labeled dataset is “extremely ambitious.” To overcome this obstacle, when building and training algorithms, scientists don’t start from scratch. They can use related techniques developed and trained for other purposes. Sometimes it is appropriate to investigate the option of developing synthetic datasets.

Machine learning algorithms for cancer prediction and treatment are often a black box. Pathologists and even researchers who developed and trained the model can’t explain how it delivers its outcome. For instance, when AI-powered software identifies the optimal cancer treatment for a particular patient, it doesn’t explain how it inferred this information. As a solution, hospitals can use explainable AI for cancer detection and treatment, where algorithms reveal the reasoning behind their decision making. This will make it easier for doctors to act upon these recommendations and explain them to patients. However, turning AI into a white box will take off some of its predictability power. So, it’s a tradeoff that healthcare organizations will have to consider making. Another consideration is about whose values the algorithms will reflect. For example, it might prioritize minimizing false positives over false negatives. If such values are not installed preemptively, algorithms might obtain them during continuous learning, and it will be difficult to tell whether they are in line with the hospitals’ policy. There are also concerns about medical data ownership and obtaining patient consent when using their information. For example, the University of Chicago shared medical records with Google to help them develop an AI-powered predictive EHR environment. As a result, both parties were slammed with a lawsuit for patient data misuse.

Healthcare organizations need to be cautious when using AI for cancer detection and prevention. Artificial intelligence is a powerful tool, which can save patients’ lives and physicians’ time. But it can also be devastating if not trained and deployed correctly. Here is a checklist that will increase your chances of success: