GUIDE-US: Grade-Informed Unpaired Distillation of Encoder Knowledge from Histopathology to Micro-UltraSound

2026-02-21

AI Summary

Current micro-ultrasound (micro-US) models struggle to accurately characterize the microscopic architecture of prostate cancer at low resolution, limiting their clinical utility for non-invasive grading. This work proposes an unpaired pathology-informed knowledge distillation approach that trains a micro-US encoder—conditioned on ISUP grade—to emulate the embedding distribution of a pretrained histopathology foundation model, without requiring patient-level image–pathology pairs or spatial registration. The method enables cancer risk stratification using ultrasound images alone, eliminating the need for pathological input during inference and thereby enhancing clinical practicality. Experimental results demonstrate improved sensitivity for clinically significant prostate cancer: a 3.5% gain at 60% specificity and a 1.2% increase in overall sensitivity compared to the current state-of-the-art.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Medical and Biological Imaging
• Search and Optimization: Distributed Search

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Purpose: Non-invasive grading of prostate cancer (PCa) from micro-ultrasound (micro-US) could expedite triage and guide biopsies toward the most aggressive regions, yet current models struggle to infer tissue micro-structure at coarse imaging resolutions. Methods: We introduce an unpaired histopathology knowledge-distillation strategy that trains a micro-US encoder to emulate the embedding distribution of a pretrained histopathology foundation model, conditioned on International Society of Urological Pathology (ISUP) grades. Training requires no patient-level pairing or image registration, and histopathology inputs are not used at inference. Results: Compared to the current state of the art, our approach increases sensitivity to clinically significant PCa (csPCa) at 60% specificity by 3.5% and improves overall sensitivity at 60% specificity by 1.2%. Conclusion: By enabling earlier and more dependable cancer risk stratification solely from imaging, our method advances clinical feasibility. Source code will be publicly released upon publication.

Problem

• prostate cancer grading
• micro-ultrasound
• histopathology
• tissue micro-structure
• non-invasive diagnosis

Innovation

• unpaired knowledge distillation
• histopathology-to-ultrasound transfer
• ISUP grade conditioning
• micro-ultrasound prostate cancer grading
• foundation model distillation