Safe AI for Prostate Cancer Diagnosis

Funding Source: NIH R21

Budget: $458,902

Time: 06/2025 - 05/2027

Safe prostate cancer segmentation and detection.

Abstract: Advances in natural language processing can build on image processing breakthroughs to offer clinicians new AI tools against prostate cancer (PCa). Current AI for interpreting mp-MRI scans relies on visual encodings like lesion annotations but fails at translation in patient care because it doesn't use the standardized PI-RADS (Prostate Imaging Reporting and Data System) format accepted by clinicians. The expertise in PI-RADS reports offers a major resource for training AI to achieve clinical acceptance. This research addresses two gaps: (1) Data availability—public PCa data repositories lack PI-RADS reports, and (2) AI modeling—existing approaches can't integrate complex radiologist expertise expressed through language. We will test the hypothesis that PI-RADS reports can be made machine-readable and combined with visual data so that AI can be trained to interpret MRIs according to the reasoning processes of radiologists. University of Delaware researchers and Memorial Sloan Kettering radiologists will collaborate to develop datasets and tools for safe AI-assisted prostate MRI interpretation.

Publications:

[ICML'25] Mengmeng Ma, Tang Li, Yunxiang Peng, Lu Lin, Volkan Beylergil, Binsheng Zhao, Oguz Akin, Xi Peng. “Why Is There a Tumor?”: Tell Me the Reason, Show Me the Evidence. In International Conference on Machine Learning, 2025. [PDF] [Code]
[CVPR'22] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, Xi Peng. Are Multimodal Transformers Robust to Missing Modality? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. [PDF] [Video]
[AAAI'21] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, Xi Peng. Multimodal learning with severely missing modality. In Proceedings of the Association for the Advancement of Artificial Intelligence, 2020. [PDF] [Code] [Video]

Open-Sourced Data:

Prostate MRI and PIRAIDl Dataset: We curate a prediction rationale dataset to tackle the lack of paired MRI images, annotations, and textual rationales for AI training. The dataset includes 180K image-mask-rationale triples with quality evaluated by expert radiologists. [https://github.com/deep-real/MedRationale].

Open-Sourced Software:

MedRationale Toolkit: Self-supervised clinical concept localization for interpretable medical image diagnosis [https://github.com/deep-real/MedRationale]