Cinepro: Robust Training of Foundation Models for Cancer Detection in Prostate Ultrasound Cineloops

TL;DR

Cinepro addresses weakly labeled cancer localization in prostate ultrasound cineloops by adapting a segmentation foundation model (MedSAM) with an involvement-aware loss aligned to pathology-reported cancer involvement and a cine-series temporal augmentation strategy. This combination leverages global image context and temporal consistency to mitigate label noise, achieving an AUROC of $77.1\%$ and a balanced accuracy of $83.8\%$ across a multi-center dataset, outperforming ROI and MIL baselines as well as standard augmentations. The approach demonstrates the feasibility of robust foundation-model training on weak ultrasound labels and its potential to improve real-time guidance during biopsies. By exploiting temporal information and strong supervision through involvement data, Cinepro provides a practical path toward more reliable prostate cancer detection in ultrasound imaging.

Abstract

Prostate cancer (PCa) detection using deep learning (DL) models has shown potential for enhancing real-time guidance during biopsies. However, prostate ultrasound images lack pixel-level cancer annotations, introducing label noise. Current approaches often focus on limited regions of interest (ROIs), disregarding anatomical context necessary for accurate diagnosis. Foundation models can overcome this limitation by analyzing entire images to capture global spatial relationships; however, they still encounter challenges stemming from the weak labels associated with coarse pathology annotations in ultrasound data. We introduce Cinepro, a novel framework that strengthens foundation models' ability to localize PCa in ultrasound cineloops. Cinepro adapts robust training by integrating the proportion of cancer tissue reported by pathology in a biopsy core into its loss function to address label noise, providing a more nuanced supervision. Additionally, it leverages temporal data across multiple frames to apply robust augmentations, enhancing the model's ability to learn stable cancer-related features. Cinepro demonstrates superior performance on a multi-center prostate ultrasound dataset, achieving an AUROC of 77.1% and a balanced accuracy of 83.8%, surpassing current benchmarks. These findings underscore Cinepro's promise in advancing foundation models for weakly labeled ultrasound data.

Figures (2)

• Figure 1: Robust training of foundation models on a multi-center dataset for PCa detection in conventional ultrasound. We train MedSAM to predict the involvement reported by pathology, and add a cine-series augmentation strategy to increase robustness.
• Figure 2: Qualitative comparison of Cinepro with our finetuned MedSAM baseline.