Images and Algorithms
Assessment of change in tumor burden is an important part of the clinical evaluation of cancer. Both tumor shrinkage and disease progression are useful endpoints in clinical trials. Evaluation of tumors is determined by radiologists using the Response Evaluation Criteria in Solid Tumors (RECIST) criteria 2.0. RECIST is a standard way to measure how well a cancer patient responds to a treatment. By using this criteria, the radiologists determine when cancer patients improve (“respond”), stay the same (“stabilize”), or worsen (“progress”) during treatment. The same tumor images are then read by an independent radiologist. Decision making in cancer treatment is subject to interobserver variability. Artificial Intelligence (AI) and machine-learning hold great promise for augmenting human radiologists, resulting in improvements in speed, cost, and precision.
The U.S. Food and Drug Administration (FDA) has found that there is a discordance of 30% between the original and confirmatory independent radiologist readings of CAT scan images from cancer clinical trials using RECIST. Project Data Sphere® (PDS), in consultation with the FDA, has initiated an Images and Algorithms program to develop machine learning algorithms that could detect, measure, and calculate total tumor burden in patients with solid tumors with a discordance rate of less than 30% by automating the confirmatory independent CAT scan reading. This automation of RECIST will yield significant cost and time savings and will benefit all stakeholders involved in the development of new oncology medicines, leading to faster delivery of innovative life-saving treatments to patients.
Modern imaging technology generates very large digital files. The information content in DICOM® (Digital Imaging and Communications in Medicine) digital language, which is the standard in clinical trials, is much richer than what is captured in the RECIST protocol. RECIST will, therefore, be superseded by AI-driven solutions that exploit the power of the digital domain.