Your Newest Colleague?
The need for AI-based end-to-end biomarkers in oncology
Jakob Nikolas Kather | | Opinion
The diagnosis of cancer is an unexpected, life-changing event for most patients – and the therapeutic path can be complex and confusing. This can be traumatic for patients and difficult for the healthcare system to manage. Complex diagnostic and therapeutic pathways cause stress, increased cost, and suboptimal outcomes (1). Addressing and reducing this burden starts with the foundation of cancer diagnostics – histopathology.
Much of a histopathologist’s day is absorbed by tedious and repetitive tasks, such as searching for tiny nests of tumor cells in prostate biopsies or counting mitotic cells in tumor tissue (2). Considering the global shortage of pathologists, an obvious step is to use AI to automate such tasks – a move that could also mitigate problems of intra- and inter-observer variability in histopathology (3). Some AI-based image analysis systems for automation of tedious routine tasks are already approved by regulatory institutions.
Although the digitization of routine pathology workflows requires significant financial and organizational investments, it enables a return on a potentially massive scale. In digitized pathology workflows, software can not only be used to view and assess images, but also to automate tedious manual processes in histopathology. However, automation of existing components in a complex diagnostic pathway can only make incremental improvements to the system as a whole. Fortunately, AI’s limits extend beyond simply automating existing steps in a diagnostic cascade. AI can be applied to complex diagnostic tasks in an end-to-end approach – all the way from raw image data to clinically actionable recommendations.
The aim of all oncology diagnostics is to improve patients’ quality of life and survival time – and, sometimes, the best way to help individuals is to improve the system. Multiple academic publications have demonstrated a direct route from AI technologies to improved diagnosis and prognosis through end-to-end AI systems. In addition to automating specific steps in the diagnostic cascade, these systems can make clinically actionable predictions directly from raw pathology images (4). For example, instead of training an AI to detect lymphocytes in cancer and use this number to forecast response to a particular immunotherapy drug, AI can be trained directly on images to learn which phenotypes are predictive of drug response (5). Without expert guidance, such end-to-end approaches usually rediscover known morphological features – but these systems can, in principle, also be “smarter” than their inventors; they can discover new morphological features in pathology slides and can detect cancer and predict subtypes better than the human eye (6,7).
In the last five years, academic research groups have pushed the frontier of AI in histopathology from basic technical studies to clinically relevant applications (2). Researchers at universities and cancer centers have shown that AI can extract a wealth of information from routine slides – far beyond what was generally assumed only five years ago (8). But, although the academic system can come up with innovative new technologies, they cannot immediately be applied to patient care. Regulatory approval for diagnostic software requires a multi-year effort and significant funding. On the other hand, the potential yield is enormous. Not long after the first academic studies were published, companies began to build products around AI technology. Currently, tech-savvy researchers and pathologists can download state-of-the-art computer code from scientific publications and apply it to their own problems. Soon, any pathologist or researcher will be able to use commercial AI solutions to extract useful information from slides. Ultimately, this technology has the potential to enable new scientific discoveries, enhance clinical trials, and improve patient outcomes in oncology and beyond.
- MJ Cordova et al., “Frequency and correlates of posttraumatic-stress-disorder-like symptoms after treatment for breast cancer,” J Consult Clin Psychol, 63, 981 (1995). PMID: 8543720.
- A Echle et al., “Deep learning in cancer pathology: a new generation of clinical biomarkers,” Br J Cancer, 124, 686 (2021). PMID: 33204028.
- DRJ Snead et al., “Validation of digital pathology imaging for primary histopathological diagnosis,” Histopathology, 68, 1063 (2016). PMID: 26409165.
- N Coudray, A Tsirigos, “Deep learning links histology, molecular signatures and prognosis in cancer,” Nat Cancer, 1, 755 (2020).
- JN Kather, J Calderaro, “Development of AI-based pathology biomarkers in gastrointestinal and liver cancer,” 17, 591 (2020). PMID: 32620817.
- R Yamashita et al., “Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study,” Lancet Oncol, 22, 132 (2021). PMID: 33387492.
- RC Maron et al., “Systematic outperformance of 112 dermatologists in multiclass skin cancer image classification by convolutional neural networks,” Eur J Cancer, 119, 57 (2019). PMID: 31419752.
- N Coudray et al., “Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning,” Nat Med, 24, 1559 (2018). PMID: 30224757.