SBIR Phase I: AI-Powered Prostate MRI Analysis Software

About This Grant

The broader impact / commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to potentially improve the accuracy, consistency, and efficiency of prostate cancer diagnosis using magnetic resonance imaging (MRI). Today, MRI scans used in prostate cancer screening can vary significantly depending on the machine and manufacturer, which makes it challenging for radiologists and software to provide consistent assessments. This project aims to create a software system that standardizes these images and provides accurate, editable 3D outlines of the prostate to support diagnostic decisions. The technology will help ensure that patients, regardless of where or how they are scanned, receive the same high-quality analysis. By reducing unnecessary imaging, streamlining radiologist workflows, and enabling more consistent evaluations, this innovation has the potential to improve early detection and reduce disparities in care. If successful, the solution will become an essential part of the prostate imaging workflow, supporting a variety of downstream clinical tools and improving access to equitable and efficient prostate cancer care. This Small Business Innovation Research (SBIR) Phase I project will develop a style-encoding generative adversarial network (GAN) to harmonize prostate MRI images from different scanner vendors and field strengths while preserving anatomical detail. A transformer-based UNETR segmentation model will then produce precise binary masks of the prostate and voxel-level uncertainty maps. The architecture includes a segmentation-aware loss function that ensures harmonized images maintain diagnostic utility when passed through a frozen segmentation model. Phase I will evaluate this pipeline using a diverse, multi-vendor prostate MRI dataset. Key performance indicators will include segmentation accuracy using the Dice similarity coefficient and Hausdorff distance, anatomical fidelity measured by structural similarity (SSIM), and radiologist editing time. Successful completion of this project will establish the technical feasibility of a scanner-agnostic, confidence-aware segmentation tool capable of supporting real-time, human-in-the-loop clinical workflows. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.