PROJECT SUMMARY (See instructions): Early detection of abdominal cancers is notoriously challenging. Symptoms are often unclear, and tumors are located deep inside the body, making regular physical checks less effective. Routine screenings aren't widely advised for those without a high-risk profile, delaying early identification. Even with advanced imaging like CT or MRI, small tumors may go unnoticed, and interpreting these scans accurately requires specialized expertise, with a risk of error. Biopsies can confirm diagnosis but are invasive and risky. Moreover, the high cost of thorough screenings and the limited reliability of tumor markers complicate the situation further. In the realm of early diagnosis, Artificial Intelligence (Al) technology, emerges as a beacon of hope, due to its power in automatically identifying key features for early diagnosis by learning from large amounts of data. However, the full potential of Al in abdominal cancer diagnostics is hindered by significant limitations, including: 1) small training data, 2) obscurity and vulnerability of Al reasoning, 2) data privacy concerns, and 3) a communication gap between medical experts and Al models. To address these challenges, this project aims at a computational framework that leverages the Al visual explanation as a channel to guide Al to reason and communicate with medical practitioners. The proposed framework will: 1) improve sample efficiency by visual explanation supervision with cancer imaging annotations, 2) consolidate Al's knowledge on reasoning across multi-institutions' data with privacy preservation, and 3) prompt Al prediction with human explanation and accelerate Al training with Al-guided explanation annotation. Our introduction of efficient, explainable Al techniques aims to reduce overdiagnosis and unnecessary monitoring, especially in high-risk groups. These technologies are crucial for assessing tumor responses to cancer treatments, offering widespread benefits. Our framework will supply open-source tools and datasets for Al researchers to test explanation quality, understand model logic, and enhance model applicability and human-Al collaboration, pushing the boundaries of artificial general intelligence. An interactive system will allow experts to easily navigate and influence Al decision-making, making this framework an excellent resource for teaching Al concepts and showing students how different aspects integrate.