Development and validation of an AI-based pathomics biomarker to predict response to first-line treatment in metastatic colorectal cancers

Microsatellite stable metastatic colorectal cancer (mCRC) patients are treated with a “one-fits-all” standard of care chemotherapy. However, responses occur in 20-30% of patients, while 15-20% are refractory. The latter are exposed to side effects that lower their quality of life. Therefore, the aim of this work is to develop a predictive biomarker, based on digital pathology images, that can help stratify patients according to their risk of resistance. Hematoxylin and eosin-stained (H&E) slides of mCRC resections were digitalized. Patches were extracted from the resulting whole slide images and automatically classified as belonging to one out of 9 classes, including the tumoral one, using a deep learning model. Based on texture features, clusters of patches were computed and were used to create the Bag of words (BoWs) that were subsequently used to train several machine learning classifiers. The best performances were obtained by a support vector machine, reaching a negative predictive value (NPV) of 90% (44/49; 95%CI=79-95%) and 82% (14/17; 95%CI=63-93%), in the training and validation sets, respectively. From a clinical perspective, NPV is the most relevant metric to ensure that sensitive patients are not wrongly prevented from receiving treatment. These preliminary findings should be further validated on a larger cohort of patients that we are collecting through a multi-institutional study.