Deep Learning Serial Radiomics for Improved Cancer Therapy Response Assessment.

Summary This Direct to Phase II SBIR proposal aims to further develop and validate a deep learning radiomics biomarker for improved early response assessment in metastatic cancer therapy trials. The motivation for this project stems from the high failure rate of oncology trials progressing from Phase II to Phase III, with 25-30% advancing and 30-40% of Phase III trials ultimately failing. The current standard for early therapy response assessment, Response Evaluation Criteria in Solid Tumors (RECIST 1.1), faces challenges, particularly in early response evaluation due to delayed tumor size changes and varied growth patterns from newer therapies. RECIST, being a categorical metric, may not fully capture early changes, as supported by previous studies that suggest more advanced metrics can provide better assessments. The proposed project addresses the need for a more sensitive and automated biomarker that can supplement RECIST in clinical trials. We propose a deep learning-based serial radiomics model designed to predict overall survival more accurately than RECIST. Our proposed model takes as input CT scans at baseline and at one or more follow up time points and outputs a continuous serial CT response score (serialCTRS) representing the probability of overall survival. An innovative aspect of our fully automated approach is that it uses the complete CT scans and therefore learns response signals both within and outside of tumor regions. A preliminary study was performed using a large real-world dataset of non small cell lung cancer patients, demonstrating that the serialCTRS biomarker improved prediction of OS compared to RECIST and compared to the percent change in tumor volume, including for therapy regimens not seen during training. SerialCTRS also demonstrated improved assessment in an external clinical trial dataset. The purpose of this project is to further develop and generalize the serialCTRS biomarker and validate its utility for informing clinical trial decisions. The project aims are: Aim 1: Expand serialCTRS by incorporating diverse datasets, self-supervised learning, multi-time point inputs, and explainability analyses; Aim 2: Validate serialCTRS performance against held-out datasets, compare to RECIST and other metrics, and demonstrate impact through simulated Phase II trials. Successful completion will lead to a validated serialCTRS tool ready for commercialization to improve clinical trial decisions, such as informing interim and futility analyses and informing the decision of whether to advance a trial to Phase III.