Artificial intelligence (AI) has been springing up in hospitals and clinics around the world in both research and direct patient care settings, with machine learning being used to predict patient outcomes, diagnose diseases, and suggest treatments. In the field of oncology, emerging AI technologies can detect tumors, diagnose cancers, and even generate chemotherapy treatment recommendations that adjust in real time on the basis of patient responses.
1. Google’s AI algorithm can detect cancer metastases with 92% accuracy. Google’s AI software encompasses a variety of healthcare functions, from predicting the amount of time a patient will spend in the hospital to their probability of being readmitted, and even assessing their risk for death. In addition to rapidly sifting through extensive medical records to assess these metrics, Google’s AI has a variety of pathologic functions. Detecting diabetic eye disease, expanding genomic research, and using digital pathology for cancer detection are among the most prominent applications.
Google’s AI cancer detection capabilities were published in a paper titled «Detecting Cancer Metastases on Gigapixel Pathology Images.» A convolutional neural network, a method that involves computers making predictions based on recognizing visual patterns, was used to detect tumors as small as 100 × 100 pixels, with an accuracy of 92.4%.[3,4] This is compared with the previous most accurate AI method, which had a tumor detection accuracy of 82.7%, whereas pathologists conducting their own manual search had an accuracy of 73.2%.
Models are trained through generating heat maps that display the probability of tumor locations, with the maximum value representing the most probable tumor location. This method reduces the false-negative rate of tumor detection by 25% compared with pathologists and by 50% compared with the previous best AI method. The majority of errors made by Google’s AI in tumor detection were related to the method of tissue preparation, primarily out-of-focus slides of tissues, which could be mitigated through more comprehensive labels for varying tissue types and improved scanning quality.
Although Google’s AI has the potential to improve the accuracy of cancer detection, further improvements in the technology are necessary in order to ensure that it is equipped for larger data sets.
2. IBM’s Watson has been implemented in more than 230 hospitals, despite allegations that its oncology system may make dangerous recommendations. IBM’s AI technology Watson is now in more than 230 hospitals across the globe, with 55 hospitals using Watson to aid in diagnosing various types of cancer. Memorial Sloan Kettering (MSK) agreed to partner with IBM and provide data for Watson, stating that «[a]s Watson Oncology’s teacher, we are advancing our mission by creating a powerful resource that will help inform treatment decisions for those who may not have access to a specialty center like MSK».
Although IBM reported that Watson can identify tumors with 93% accuracy, many physicians are not convinced of its capabilities. An instance in which Watson suggested that a 65-year-old patient with lung cancer and severe bleeding receive bevacizumab despite the drug’s associations with severe hemorrhaging caused many doctors to doubt Watson’s accuracy. Although many of these doctors believed Watson was using real patient data to generate these recommendations, they discovered that Watson was using hypothetical data and simply relaying recommendations from other doctors rather than synthesizing its own conclusions from large data sets. IBM officials responded that Watson is constantly learning, and the 11 software developments seen in the past year are a testament to the impact of physician criticism and feedback.
Despite the controversial recommendations, many doctors continue to use Watson, most often when there is not consensus on how to treat a patient. Research on Watson conducted in India was presented at the American Society of Clinical Oncology meeting last June and found that recommendations made by Watson were the same recommendations made by physicians 81% of the time with colon cancer, 93% of the time with rectal cancer, and 96% of the time with lung cancer. In addition, a study in Thailand demonstrated that Watson can screen for lung cancer and breast cancer 78% faster than a pathologist can.
Because Watson gathers data from the likes of PubMed, the Wellcome Sanger Institute’s Catalogue of Somatic Mutations in Cancer database, and the National Cancer Institute’s drug dictionary, and can read 800 million pages per second, its main appeal in healthcare is the ability to keep physicians up to date with new research at a rate that wouldn’t be possible manually.
3. Deep learning can improve the accuracy of pathologists’ cancer diagnoses. The 2016 Camelyon Grand Challenge, a competition on cancer metastasis detection in lymph node images that is hosted by the International Symposium on Biomedical Imaging, showcased the dramatic results of deep learning in AI technology. Harvard Medical School’s Beth Israel Deaconess Medical Center used deep learning to drop the error rate of human diagnosis by 85%.
Similar to Google’s AI technology, the Beth Israel team fed data into a system that sorted tumor location into probability heat maps. Through this deep learning mechanism, the computer was able to identify cancer with 92% accuracy. Although Beth Israel’s AI software cannot compare with the 96% accuracy of human pathologists, the team showed that pairing the diagnoses of pathologists with AI yields favorable results.
Since the competition, the team has released data that displayed 97.1% accuracy, thereby surpassing the accuracy of pathologist’s diagnoses alone. Combining deep learning AI with a pathologist has yielded 99.5% accuracy.
Deep learning has a range of other applications beyond tumor detection. Padideh Danaee and colleagues from Oregon State University used deep learning in order to extract meaningful information from gene expression data. Genes that are crucial for breast cancer diagnosis were extracted from gene expression data, and interactive genes with the potential to serve as cancer biomarkers were highlighted. Although many current algorithms can distinguish normal cells from abnormal and cancer cells, Danaee and her team identified that there is limited technology that can identify the genes that play a critical role in cancer. Stacked denoising autoencoders were used, which convert noisy gene expression data into more concise and meaningful data in order to be used by AI. These data were then sorted into healthy control samples and breast cancer samples, which were subsequently broken down into significant genes and relevant biomarkers in cancer.
Deep learning has the ability to build upon these data and recognize patterns, which illustrates its potential role in cancer detection. However, like other AI companies and research teams, Danaee and colleagues emphasized the limitations that come with the unavailability of large data sets.
4. Artificial intelligence can pinpoint genetic mutations. Although many AI companies are only beginning to move into healthcare, the ultimate objective is eventual commercialization of AI technology. One company with a head start is Sophia Genetics, a Switzerland-based AI healthcare company that is currently being used in more than 420 hospitals. By rapidly identifying genetic mutations, Sophia Genetics hopes to expand genomic research and «democratize data-driven medicine.» The technology costs anywhere from $50 to $200 per genomic test and can detect, annotate, and preclassify genomic variants not only in the field of oncology, but in cardiology, pediatrics, and other specialties.
Another company with a genomics-based platform is Freenome, which uses AI to identify biomarkers in blood, alterations in gene expression, fluctuations in immune system activity, and cancer-related proteins. Using a database of these biomarkers, Freenome is developing a range of noninvasive blood tests to not only detect cancer, but also recommend treatments.
5. A self-learning AI system developed at MIT can observe treatment regimens and adjust dosages in real time to generate optimal treatment plans. In an attempt to mitigate the debilitating side effects of radiation and chemotherapy, a team at MIT has developed a model that can reduce the toxicity of cancer treatments through a self-learning machine. A 50-patient simulated trial was conducted, which demonstrated that the self-learning machine could potentially reduce the toxicity of these cancer treatments by up to 50% while maintaining the same effect with regard to tumor shrinkage. Principal investigator Pratik Shah emphasized how they maintain the same goal of reducing tumor sizes but hope to increase patients’ quality of life by reducing harmful side effects.
In one trial simulation, the model was adapted for glioblastoma using traditional chemotherapy regimens including vincristine, temozolomide, lomustine, and procarbazine. Initially, the AI reads through established cancer treatment methods from clinical trials. Then, over several weeks or months, the self-learning machine adjusts the dosages of all four of these drugs in response to the mean diameter of the tumor and the toxicity of the drugs. Because the AI runs off of a reinforced learning model where there are rewards and punishments for certain behaviors, the machine is penalized if it decides to administer a full dosage. Over 20,000 trial-and-error test runs were constructed by the model after the 50-person simulated trial, allowing it to use those parameters when forming new treatment methods. In addition, because the model is specifically designed around the patient it is observing, the AI can consider such factors as medical histories, genetic profiles, and various biomarkers, whereas clinical trials of large populations often cannot.
- Bergen M. Google is training machines to predict when a patient will die. Bloomberg. June 18, 2018. Source Accessed August 13, 2018.
- Healthcare and biosciences. Google AI. Source Accessed August 13, 2018.
- Liu Y, Gadepalli K, Norouzi M, et al. Detecting cancer metastases on gigapixel pathology images. Cornell University Library. March 8, 2017. Source Accessed August 13, 2018.
- Cornelisse D. An intuitive guide to convolutional neural networks. freeCodeCamp. April 24, 2018. Source Accessed August 13, 2018.
- Palmer A. IBM’s Watson AI suggested ‘often inaccurate’ and ‘unsafe’ treatment recommendations for cancer patients, internal documents show. Daily Mail. July 27, 2018. Source Accessed August 13, 2018.
- Somashekhar SP, Sepúlveda MJ, Norden AD, et al. Early experience with IBM Watson for Oncology (WFO) cognitive computing system for lung and colorectal cancer treatment. Program and abstracts of the 2017 American Society of Clinical Oncology Annual Meeting; June 2-6, 2017; Chicago, Illinois. Source Accessed August 13, 2018.
- Watson Oncology. Memorial Sloan Kettering Cancer Center. Source Accessed August 13, 2018.
- Schmidt C. M.D. Anderson breaks with IBM Watson, raising questions about artificial intelligence in oncology. J Natl Cancer Inst. 2017;109.
- Kontzer T. Deep learning drops error rate for breast cancer diagnoses by 85%. Nvidia. September 19, 2016. Source Accessed August 13, 2018.
- Wang D, Khosla A, Gargeya R, Irshad H, Beck AH. Deep learning for identifying metastatic breast cancer. Cornell University Library. June 18, 2016. Source Accessed August 13, 2018.
- Danaee P, Ghaeini R, Hendrix D. A deep learning approach for cancer detection and relevant gene identification. Pac Symp Biocomput. 2017;22:219-229.
- Cynober T. Artificial intelligence in oncology: fantasy or reality? Labiotech. April 7, 2018. Source Accessed August 13, 2018.
- The universal technology for genomic analysis. Sophia Genetics. 2018. Source Accessed August 13, 2018.
- Using AI to detect cancer. Freenome. Source Accessed August 13, 2018.
- Institut Curie and Freenome announce a strategic collaboration in cell-free DNA analysis using machine learning technology. Business Wire. March 29, 2018. Source Accessed August 13, 2018.
- AI model ‘learns’ from patient data to make cancer treatment less toxic. Science Daily. August 10, 2018. Source Accessed August 13, 2018.
- Matheson R. Artificial intelligence model «learns» from patient data to make cancer treatment less toxic. MIT News Office. August 9, 2018. Source Accessed August 13, 2018.
A Product Manager with expertise in pharma marketing and sales operations