Summary
Prostate cancer diagnosis can rely on both tissue scans under a microscope and ultrasound images, but matching these two data types is tricky. The authors devise a way to better teach computers to find cancerous tissue in ultrasound images by using information about how much cancer a tissue sample contains. This method helps the computer focus on specific parts of the ultrasound images, improving its ability to distinguish cancer grades more accurately than previous methods. The approach was tested on thousands of biopsy samples from hundreds of patients and showed notable improvements in detecting significant cancers.
What this means in practice
- •For medical imaging developers: Develop improved prostate cancer ultrasound analysis tools that better identify malignant tissue using weakly supervised alignment with pathology data.
- •For biopsy image analysis teams: Enhance cancer grading software to focus attention on the most relevant ultrasound tissue patches guided by biopsy involvement percentages.
Authors
Obed Korshie Dzikunu, Emma Willis, Mohammad Mahdi Abootorabi, Mohamed Harmanani, Zhuoxin Guo, Ferdinand Luger, Adam Kinnaird, Brian Wodlinger, Parvin Mousavi, Purang Abolmaesumi
Abstract
Unpaired cross-modal distillation transfers grade structure from histopathology into a micro-ultrasound (micro-US) encoder by aligning a pooled needle-region embedding to a frozen histopathology teacher under grade-group correspondence alone. A single objective is thereby required to serve two distinct functions: rendering patch features discriminative of tissue state, and selecting which patches enter the pooled representation. We decouple them. Weak spatial supervision derived from percentage involvement, recorded routinely at biopsy, constrains the predicted proportion of malignant tissue within each core, acting on the encoder features independently of the alignment objective. The alignment loss then operates on features that differ across a core, and attention concentrates on a subset of patches rather than remaining near-uniform. On 7,166 biopsy cores from 811 patients across seven centers under patient-level 5-fold cross-validation, the method reaches 67.1 macro AUC and 68.5 csPCa AUC, against 61.2 and 52.8 for the existing unpaired alignment method and 63.1 and 62.6 for the strongest unimodal baselines. Ablation against existing attention regularizers designed to prevent attention-uniformity collapse shows that such regularizers do not substitute for label-derived supervision: they constrain the attention distribution, whereas the signal required acts on the features that attention reads.