Learning Prostate Anatomy at Test Time for Cancer Detection in Micro-Ultrasound

Obed Korshie Dzikunu; Mohammad Mahdi Abootorabi; Mohamed Harmanani; Paul F. R. Wilson; Emma Willis; Ferdinand Luger; Adam Kinnaird; Brian Wodlinger

cs.CVAug 20, 2026

Learning Prostate Anatomy at Test Time for Cancer Detection in Micro-Ultrasound

Authors: Obed Korshie Dzikunu, Mohammad Mahdi Abootorabi, Mohamed Harmanani, Paul F. R. Wilson, Emma Willis, Ferdinand Luger, Adam Kinnaird, Brian Wodlinger, +2 more

Organizations: Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada · Vector Institute, Toronto, Canada · School of Computing, Queen’s University, Kingston, Canada · Ordensklinikum Linz, Linz, Austria · Department of Surgery, University of Alberta, Edmonton, Canada · Exact Imaging, Markham, Canada

Abstract

Domain shift across clinical centers using different imaging hardware or acquisition protocols remains a fundamental barrier to deploying deep learning models for prostate cancer (PCa) detection. Existing test-time adaptation (TTA) methods address distribution shift through entropy minimization or augmentation-based self-supervision, correcting for statistical differences in image appearance but ignoring the anatomical structure of the target domain. We propose ANT, a segmentation-guided TTA framework that adapts a pretrained cancer detection encoder to the target domain by solving an auxiliary prostate segmentation task at test time, supervised by pseudo-masks from a frozen pretrained segmentation network. By aligning encoder representations to prostate anatomy in the target domain, ANT corrects domain-specific feature drift while preserving cancer-discriminative structure. The model was trained on 693 patients imaged with an earlier-generation micro-ultrasound scanner in a multi-center clinical trial, and evaluated on 118 patients acquired with a newer-generation system across two centers in another clinical trial. Under a leave-one-center-out protocol with identical evaluation conditions across all methods, ANT improves mean AUC by 2.9% and 3.6% at the biopsy-core and patient levels, respectively, over no adaptation, outperforming TTA baselines. Code is available at: https://github.com/ObedDzik/ant.git.

Figures & tables

Figure 1: Overview of ANT framework. At inference, the first $n$ layers of the trained cancer detection encoder is domain-adapted via an auxiliary segmentation task guided by pseudo-masks from a frozen segmentation network. The adapted encoder is then used for cancer detection.

Center	Patients	Cores	Cancer %
A (source)	$693$	$5479$	$16\%$
B	$76$	$1118$	$34.6\%$
C	$42$	$569$	$27.0\%$

Table 1: Cohort statistics per clinical center. Centers differ in scanner generation and acquisition protocol, representing realistic deployment-time domain shift.

	Patient-Level csPCa AUC		Core-Level csPCa AUC		Core-Level PCa AUC
Method	B	C	B	C	B	C
No adaptation	$77.6(72.9-82.0)$	$79.4(66.8-89.9)$	$78.5(73.6-83.3)$	$84.5(74.0-93.1)$	$71.5(68.2-74.6)$	73.8 (68.5 – 79.3)
MEMO ( Zhang et al., 2022 )	$78.3(73.8-82.5)$	$83.5(73.0-91.7)$	$77.7(73.1-82.5)$	$83.4(71.8-92.9)$	$73.3(70.2-76.3)$	$70.6(65.3-76.3)$
TENT ( Wang et al., 2020 )	$76.0(71.6-80.2)$	$75.0(60.8-87.0)$	$80.1(76.0-84.3)$	$81.9(70.3-92.6)$	$74.5(71.5-77.3)$	$68.6(62.7-74.4)$
EATA ( Niu et al., 2022 )	$73.5(68.7-77.2)$	$69.4(65.1-72.8)$	$79.8(75.3-82.4)$	$80.5(77.3-84.2)$	74.9 (71.0 – 77.2)	$71.9(69.8-74.5)$
CoTTA ( Wang et al., 2022 )	$77.9(73.2-82.2)$	$77.9(64.6-89.7)$	$78.5(73.7-83.3)$	$84.1(73.8-92.7)$	$71.4(68.2-74.5)$	$72.9(67.4-78.4)$

Table 2: AUC (95% bootstrap CI) for all methods across both evaluation centres (B and C). Bold indicates the best performance per column.

	Sensitivity @ 60% Specificity (%)		Sensitivity @ 80% Specificity (%)
Method	B	C	B	C
No adaptation	$81.0(73.5-88.2)$	$85.3(43.7-100.0)$	$67.1(56.1-76.5)$	$70.1(43.7-94.1)$
MEMO ( Zhang et al., 2022 )	$82.5(75.3-89.4)$	$85.2(64.7-100.0)$	$61.7(51.3-72.1)$	$74.5(50.0-95.6)$
TENT ( Wang et al., 2020 )	$80.4(72.7-87.5)$	$83.1(47.1-100.0)$	$66.4(57.8-75.4)$	$71.1(47.1-91.7)$
EATA ( Niu et al., 2022 )	$80.2(71.3-86.4)$	$89.5(49.2-100.0)$	$65.5(57.3-73.2)$	$63.2(44.6-88.5)$
CoTTA ( Wang et al., 2022 )	$80.0(72.0-87.3)$	$83.4(53.3-100.0)$	$65.8(55.8-75.5)$	$75.2(53.3-94.4)$

Table 3: Sensitivity at fixed specificity (95% bootstrap CI) for all methods across both evaluation centres (B and C) at the core level for clinically significant prostate cancer (csPCa). Bold indicates the best performance per column.

Dice Tertile	$n$	Dice Median (Range)	ANT	Baseline	ANT $>$ Base	ANT	Baseline	ANT $>$ Base
			PCa AUC (%)			csPCa AUC (%)
Low $(\leq 0.72)$	11	$0.60(0.36-0.71)$	$80.0(29.6-100.0)$	$78.6(23.3-100.0)$	4/11	$100.0(78.1-100.0)$	$80.0(65.6-82.7)$	3/3
Mid $(0.72-0.90)$	10	$0.83(0.72-0.90)$	$68.3(6.2-100.0)$	$76.2(18.2-100.0)$	3/10	$63.6(50.0-95.8)$	$54.5(36.7-70.8)$	5/5
High $(>0.90)$	11	$0.95(0.91-0.99)$	$57.1(21.4-94.4)$	$44.0(21.4-96.3)$	6/11	$71.7(40.0-100.0)$	$65.7(23.1-100.0)$	4/6
Overall	32	$0.83(0.36-0.99)$	-	-	13/32	-	-	12/14

Table 4: Per-patient AUC for ANT and the unadapted baseline stratified by MicroSegNet pseudo-mask Dice score tertile. Values reported as median (range). csPCa AUC is available only for patients with at least one positive and one negative core for clinically significant cancer.

	Patient-Level csPCa AUC		Core-Level csPCa AUC		Core-Level PCa AUC
Method	Center B	Center C	Center B	Center C	Center B	Center C
PNF+ ( Wilson et al., 2026 )	$71.2(66.8-75.6)$	$75.6(64.4-85.3)$	$72.4(68.2-76.5)$	$75.4(62.8-87.2)$	$69.8(66.6-73.1)$	$63.8(58.1-69.4)$
PNF+ + ANT (ours)	72.8 (68.3 – 77.4)	80.5 (69.7 – 88.8)	73.9 (69.8 – 78.0)	78.2 (66.5 – 87.1)	∗ 73.1 (70.4 – 76.3)	65.2 (59.3 – 70.5)

Table 5: ANT applied to ProstNFound+ (PNF+), the current state-of-the-art model for micro-ultrasound prostate cancer detection. Results (AUC,

\%

) demonstrate model-independent generalization of the proposed TTA approach. Best results in bold.

	Sensitivity @ 60% Specificity (%)		Sensitivity @ 80% Specificity (%)
Method	B	C	B	C
PNF+ ( Wilson et al., 2026 )	$76.4(68.6-84.3)$	$72.3(50.0-92.9)$	$44.5(33.3-56.0)$	58.4 (30.4 – 87.5)
PNF+ + ANT (ours)	78.9 (70.6 – 86.4)	85.5 (65.0 – 100.0)	46.0 (36.2 – 56.7)	$56.0(30.8-77.8)$

Table 6: Sensitivity at fixed specificity (95% bootstrap CI) for the unadapted ProstNFound baseline (PNF) and ANT across both evaluation centres (B and C) at the core level for clinically significant prostate cancer (csPCa).

Figure 2: Left: AUROC for clinically significant cancer prediction evaluated across increasing involvement thresholds (cores with true involvement $\geq t$ or benign). Right: Distribution of activation leakage, defined as the ratio between average prostate heatmap activation and average needle-region activation, stratified by cancer status (benign vs PCa) and method, PNF+ (no TTA) vs. PNF+ + ANT. Lower values indicate better spatial specificity, i.e., reduced off-target activation outside the biopsy needle region.

Figure 3: Representative heatmaps of model activations without TTA (PNF+) and with TTA (PNF+ + ANT) on the same biopsy cores. Activations represent predicted involvement, with higher values indicating greater likelihood of clinically significant cancer. Each example is annotated with grade group (GG) and corresponding involvement percentage.

Layer position and parameter type
Update target	Position	Patient-Level (csPCa)	Core-Level (csPCa)	Core-Level (PCa)
Full blocks	First $n$	82.1	84.4	72.7
Full blocks	Middle $n$	$79.8$	$81.9$	73.6
Full blocks	Last $n$	$77.2$	$80.8$	$72.7$
Full blocks	All blocks	$74.9$	$79.6$	$70.2$
Norm only	First $n$	$76.4$	$80.9$	$72.8$

Table 7: Ablation on encoder update strategy and auxiliary task. Mean AUC (

\%

) across held-out test centers. Best in bold.

Inner steps $\tau$	Patient-Level (csPCa)	Core-Level (csPCa)	Core-Level (PCa)
0(w/o ANT)	$78.5$	$81.5$	$72.7$
2	$76.0$	$81.1$	$72.6$
5	$78.9$	$83.4$	$72.5$
10	82.1	84.4	72.7
20	$78.2$	$78.0$	$72.0$

Table 8: Ablation on number of TTA inner steps. Mean AUC (

\%

) across held-out test centers. Best in bold.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
TTA learning rate ( $\eta_{\mathrm{TTT}}$ )	$10^{-4}$
Encoder learning rate ( $\eta_{\mathrm{enc}}$ )	$10^{-5}$
Inner steps ( $\tau$ )	10
Blocks to adapt ( $n$ , DINOv3)	6
Blocks to adapt ( $n$ , MedSAM)	3
Segmentation head hidden dim	256

Appendix

Table 9: Selected TTA hyperparameters.

Method	Time (s)
No adaptation	$0.07\pm 0.01$
TENT	$0.11\pm 0.03$
EATA	$0.13\pm 0.01$
ROID	$0.19\pm 0.03$
SAR	$0.22\pm 0.02$
CoTTA	$0.26\pm 0.02$

Appendix

Table 10: Mean wall-clock inference time per image (seconds

\pm

std) on a single NVIDIA L40S GPU (48 GB). ANT includes

\tau=10

inner optimisation steps per image.

	Patient-Level csPCa AUC (%)		Core-Level csPCa AUC (%)		Core-Level PCa AUC (%)
Ordering	B	C	B	C	B	C
Fixed order	$79.9\pm 0.03$	$84.1\pm 0.08$	$81.1\pm 0.02$	$88.0\pm 0.06$	$73.1\pm 0.03$	$72.3\pm 0.03$
Random order	$80.0\pm 0.04$	$82.3\pm 0.03$	$80.6\pm 0.04$	$85.8\pm 0.08$	$72.9\pm 0.01$	$72.0\pm 0.05$

Appendix

Table 11: Ordering sensitivity analysis. Mean AUC (%)

\pm

std across five random seeds for fixed and randomly permuted test set orderings.

Subgroup (Involvement)	$n$	Baseline ECE (%)	ANT ECE (%)	$\Delta$ ECE (%)
Overall	1,118	7.6	2.7	$-4.9$
Low	161	29.7	32.0	$+2.2$
Mid	93	19.6	16.5	$-3.0$
High	131	14.3	13.8	$-0.5$

Appendix

Table 12: Calibration analysis on Center B (

n=1{,}118

cores). Expected Calibration Error (ECE) is reported for the unadapted baseline (PNF+) and ANT.

\Delta

ECE = ANT ECE

-

Baseline ECE; negative values indicate improved calibration.

Explore similar work

Jun 29, 2026

Learning Where to Look: A Reinforcement Learning Framework for Robust Micro-Ultrasound Prostate Cancer Detection

Micro-ultrasound (

μ

US) is a new, emerging, and promising imaging modality for prostate cancer (PCa) detection, but accurate identification of suspicious tissue remains highly dependent on clinical experience, leading to substantial inter-observer variability. Machine-learning assistance can reduce this variability; however, training reliable deep models is challenging because supervision is sparse and noisy -- typically limited to core-level histopathology outcomes (e.g., cancer grade and its percentage in a biopsy core) without pixel-level lesion annotations and under severe class imbalance. We introduce Prost-RL, which reframes

μ

US PCa detection as a spatially aware, policy-driven inference problem by learning where to look before decoding. Prost-RL integrates a lightweight reinforcement-learning policy into a foundation-model encoder-decoder to generate interpretable spatial attention maps that act as soft prompts for both cancer-likelihood heatmap prediction and image-level classification. We further propose Adaptive Policy Optimization (APO) to stabilize hybrid supervised-RL training and a noise-robust objective combining symmetric cross-entropy with negative-entropy regularization to mitigate weak-label noise and encourage sharp localization. On a cohort of 6,607 biopsy cores from 693 patients across five clinical sites, Prost-RL achieves

79.0\pm3.5

AUROC with

64.6\pm6.3

% sensitivity at 80% specificity for core-level detection (+2.1 AUROC and +4.5 sensitivity points over the strongest baseline), and

79.3\pm5.8

AUROC for clinically significant cancer classification. The learned policy highlights biopsy-aligned regions, providing transparent, spatially grounded evidence alongside quantitative risk predictions. Code is available at: https://github.com/DeepRCL/Prost-RL.

Mohammad Mahdi Abootorabi, Sina Namazi, Armin Saadat +7

The University of British Columbia, Vancouver, BC, Canada · Vector Institute, Toronto, ON, Canada · Queen’s University, Kingston, ON, Canada +1

Medical Imaging Reinforcement Learning Medical Image Classification Ultrasound Imaging Weakly Supervised Learning Oncology+1

Ask PDF

Aug 14, 2026

Zero-Shot Adaptation of Medical Vision Foundation Models for High-Frequency Micro-Ultrasound Prostate Segmentation

Prostate cancer claims a life every 80 seconds. Early detection is needed to prevent disease progression, and both PSA density calculation and biopsy decisions rely on knowing the exact boundary of the gland. Conventional ultrasound at 6-12 MHz blurs this boundary, missing one in three high-risk cancers. Micro-ultrasound (29 MHz) improves resolution threefold but introduces dense acoustic speckle that obscures the outer wall; given the same image, two clinicians draw outlines differing by over 10% in area. Supervised methods are costly and generalise poorly across scanners. Can a foundation model segment the prostate with no training data? We present the first zero-shot pipeline for this modality: MedSAM, pre-trained on over 1.5 million medical images, localises the prostate; we then apply CLAHE to sharpen the outer wall, binary dilation to recover missed pixels, and Fourier smoothing (4 modes, s=1.05) to refine the boundary. MedSAM requires a spatial prompt, so we evaluate bounding-box and point-click strategies across 75 patients of the Micro-Ultrasound Prostate Segmentation dataset (2,621 slices). On the 20-patient held-out test set, the pipeline reduces mean boundary-distance error by 45% (Dice 0.749+/-0.043 to 0.865+/-0.029; HD95 217.2+/-36.9 to 120.1+/-26.1 px), reaching Dice 0.859 across the cohort. Its mean overlap shows no significant difference from the three non-expert rater groups (p>0.19), while segmenting 38-52% more consistently (lower inter-patient standard deviation). Point-click prompts fail regardless of placement (best Dice=0.350), because speckle gives no stable local contrast. Only an approximate bounding box is required, so any clinic can deploy it without data collection, annotation, or retraining.

Ayusha Abbas, Saram Abbas, Kabita Adhikari

Newcastle University Newcastle Upon Tyne, UK

Ultrasound Image Segmentation Medical Image Segmentation Zero-Shot Medical Image Segmentation Boundary-Aware Segmentation Medical Imaging Foundation Models

Ask PDF

Jul 8, 2026

Compass: Prostate Cancer Detection Needs Multi-View Context

Artificial intelligence (AI) analysis of micro-ultrasound (

μ

US) has shown promise for prostate cancer (PCa) detection. However, most existing AI methods focus on the analysis of single

μ

US images in isolation. By contrast, expert

μ

US readers typically assess a full recorded video study, which provides three-dimensional context, to improve PCa detection compared to single-frame analysis. Inspired by this clinical workflow, we propose Compass, a novel AI methodology which models a

μ

US study as a stream of 2D images. Compass jointly integrates rotational sweep videos of the prostate with

μ

US frames acquired at the moment of biopsy, and performs evidence aggregation across the study using a transformer conditioned on the probe's rotational angle. Finally, a decoder head predicts frame-level and study-level risk scores for the patient. The model is trained and evaluated using a multi-center clinical trial dataset of

μ

US studies, including continuous rotational scans of the prostate and videos captured during biopsy acquisition. We compare the proposed method to baseline AI methods from the literature and to risk scores provided by clinical experts. Our framework shows strong performance, highlighting the value of multi-view context for

μ

US PCa detection, and providing a potentially powerful tool to complement human expertise in

μ

US-based PCa diagnosis. Our code is available at: https://github.com/mharmanani/Compass.

Paul F. R. Wilson, Mohamed Harmanani, Zhuoxin Guo +6

Queen’s University, Kingston, Canada · Vector Institute, Toronto, Canada · University of British Columbia, Vancouver, Canada +3

Medical Imaging Multi-View Learning Medical Image Classification Ultrasound Imaging Oncology

Ask PDF