Keywords: tumors, deep learning, AI, surveillance, early detection, healthcare
Summary:Brain tumor surveillance is a prevalent practice in neuro-oncology that aims at detecting signs of cancer progression. The status quo in the clinic consists of radiologists relying on the visual inspection and on computing the bi-products of the largest perpendicular diameters from a single 2-dimensional axial image to determine if a tumor had grown or not. If radiation therapy is necessary, radiation therapists and physicists perform manual tumor contouring, also referred to as “delineation” or “segmentation”, to optimize treatment. This practice is highly prone to inter-observer variability and associated with delays in tumor growth detection – up to three years and an increase of more than 170% in tumor volume. Our recent work established the importance of longitudinal analysis of tumor volumes in assisting physicians to diagnose and monitor tumor growth at significantly earlier times (2-3 years earlier) and at significantly smaller tumor volumes than the current standard method. This underscores the need for an objective and reproducible auto-contouring method for brain tumor segmentation and volumetric measurements that will reform tumor surveillance in radiology and improve consistency of radiotherapy. Recently, a plethora of Artificial Intelligent (AI) systems, based on deep learning (DL) algorithms, have been proposed for the analysis, i.e., classification and segmentation, of radiological images. However, robustness and trustworthiness are typically overlooked in these designs. Perhaps more unsettling is the discovery of adversarial attacks, which exposed the vulnerabilities of DL systems and raised concerns about their practical deployment. The mechanisms of adversarial attacks are still unknown due to the complexity of the DL models. While much of the interest with adversarial attacks has stemmed from their ability to shed light on possible limitations of current DL methods, most notably the cybersecurity threats they pose, there is a disengagement and under-exploration of widespread AI automation of medical applications given adversarial vulnerabilities. We are pioneering DL frameworks that include an awareness of their limitations and have the capacity for insightful introspection by assessing their performance and predicting breakdown through uncertainty propagation during training. This project advances a trustworthy AI-based tumor segmentation method that provides a confidence map of the segmentation and is robust to artifacts and adversarial attacks. High accuracy consolidated with high confidence estimates will establish the trustworthiness of AI-aided systems and move towards meaningful integration into the clinical practice. The overall solution is based on the physician-in-the-loop (PIL) philosophy, where the physician can quickly review and adjust segmentation decisions based on uncertainty information. The adjustments made by the physician will be subsequently used to retrain and improve the AI system. Our team has filed two patent applications to protect our IP.