GUIDE-US: Grade-Informed Unpaired Distillation of Encoder Knowledge from Histopathology to Micro-UltraSound
Authors: Emma Willis, Tarek Elghareb, Paul F. R. Wilson, Minh Nguyen Nhat To, Mohammad Mahdi Abootorabi, Amoon Jamzad, Brian Wodlinger, Parvin Mousavi, +1 more
Organizations: Queen’s University, Kingston, Canada. · University of British Columbia, Vancouver, Canada∗. · Exact Imaging, Markham, Canada.
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.
Figures & tables
Benign
ISUP 1
ISUP 2
ISUP 3
ISUP 4
ISUP 5
Total
5727
0
480
195
134
71
Table 1: Micro-US ISUP grade-group distribution (number of cores).
Benign
ISUP 1
ISUP 2
ISUP 3
ISUP 4
ISUP 5
Train
2083
1919
967
896
900
882
Val
260
240
121
111
112
110
Test
260
240
121
111
112
110
Table 2: PANDA ISUP distribution by split (number of WSIs).
Figure 1: Overview of the two-stage pathology-guided teacher-student framework for micro-US ISUP grading. In Stage 1 (top), a histopathology teacher model is trained on whole-slide images (WSIs) using attention-based multiple-instance learning (MIL) for aggregated patch embeddings and a multi-layer perceptron (MLP) for ISUP classification, with weights subsequently frozen. In Stage 2 (bottom), the micro-US student model undergoes multi-task training: an image encoder extracts the full image patches features, then the biopsy-region patches are aggregated via attention-based MIL into embeddings that are guided to match frozen pathology representations through contrastive knowledge distillation. Concurrently, an image decoder generates full-image cancer masks, enabling ROI isolation for weakly supervised segmentation guided by coarse histopathology reports.
Sens@40%Spec
Sens@60%Spec
Sens@80%Spec
Method
AUROC
Entropy
PCa
csPCa
PCa
csPCa
PCa
csPCa
SAM Kirillov et al. (2023)
69.2
8.4
76.2
73.1
67.0
65.4
52.5
44.6
MedSAM Kirillov et al. (2023)
73.5
8.1
80.8
79.9
71.6
72.2
57.1
51.4
ProTeUS Elghareb and Harmanani (2025)
76.0
7.4
85.4
87.2
76.2
79.5
61.7
58.7
ProstNFound+ Wilson et al. (2025)
77.9
7.32
89.4
92.1
79.1
82.5
64.5
60.2
Ours
78.4
6.78
89.3
92.6
80.3
86.0
64.7
64.3
Table 3: Quantitative results for micro-US models. AUROC, Entropy, and Sensitivity at multiple specificity operating points are reported as the mean across 5-fold cross-validation.
Figure 2: AUROC per ISUP grade group for biopsy-region features from the encoder, comparing our method (orange) to the ProstNFound+ baseline (blue).
Figure 3: Ablation studies. (a) Triplet loss vs. CLIP-style contrastive loss across key metrics. (b) Step-wise ablation of proposed components showing cumulative gain in Sens@60%Spec (csPCa) from the ProstNFound+ baseline. (c) Sensitivity to triplet-loss weight α (mean across folds) on Sens@60%Spec (csPCa).
Image Encoder
AUROC
Entropy
Sens@60%Spec
Sens@60%Spec (csPCa)
UNI-2
78.6
6.75
81.2
87.1
GigaPath
78.0
6.5
82.0
87.7
HistoSSLScaling
78.2
6.8
80.5
85.6
Table 4: Teacher robustness: swapping the pathology teacher encoder (same tiles, different feature extractors).
Figure 4: Qualitative comparison of cancer probability heatmaps between our method (top row) and the ProstNFound+ baseline (bottom row) on micro-US cases. The biopsy regions are overlaid on each image. The ground-truth ISUP grade and cancer involvement for each case are shown.
Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada · Vector Institute, Toronto, Canada · School of Computing, Queen’s University, Kingston, Canada +3