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, Parvin Mousavi, Purang AbolmaesumiarXiv 2026 · arXiv

TL;DR

The authors introduce ANT (Anatomical Network Adaptation at Test-time), a novel test-time adaptation framework that combats domain shift in prostate cancer detection from micro-ultrasound videos. By using a frozen, pretrained segmentation network to guide the early layers of a cancer detection encoder at inference time, ANT anchors feature representations to domain-invariant anatomy, significantly outperforming standard entropy-based adaptation baselines.

What Problem Does This Paper Address?

Deep learning models for prostate cancer (PCa) detection often suffer severe performance drops when deployed at new clinical sites or on newer generations of imaging hardware—a challenge known as domain shift.

While Test-Time Adaptation (TTA) offers a way to adjust model parameters on-the-fly using unlabeled deployment data, existing methods typically rely on generic statistical signals like entropy minimization or data augmentation consistency. These methods modify global feature statistics but completely ignore the underlying anatomical structure of medical images. Under large inter-center or inter-device shifts, purely statistical TTA can degrade performance or introduce spurious, off-target predictions.

Methods

The authors propose ANT, a segmentation-guided TTA framework. The core insight is that while low-level image statistics change across scanner generations, the macro-anatomy—the prostate gland itself—remains a stable, identifiable structure.

Overview of the ANT framework showing the training phase and the test-time anatomical adaptation process.

The adaptation process works as follows during inference:

  1. Pseudo-Mask Generation: For each incoming target image, a frozen, pretrained generalist segmentation model (MicroSegNet) generates a prostate pseudo-mask.
  2. Auxiliary Task Optimization: A lightweight, randomly initialized segmentation head is attached to the model's image encoder. The model minimizes a combined Dice and Binary Cross-Entropy loss to align its internal representations with the pseudo-mask.
  3. Selective Layer Updates: Crucially, only the first $n$ blocks of the Vision Transformer (ViT) encoder are updated. These early layers process low-level signal statistics where hardware-induced domain shifts are most prominent.
  4. Episodic Reset: To prevent cross-sample leakage, the encoder is reset to its baseline state after making a prediction for a sample, while the lightweight segmentation head persists across the test set to accumulate a refined understanding of the target domain's anatomy.

Datasets

The framework was trained and evaluated using data from two distinct multi-center clinical trials:

  • Source Domain (Training): Retrospective data from 693 patients (5,479 biopsy cores) collected between 2013 and 2016 across five institutions (collectively called Center A). Images were acquired using an earlier-generation ExactVu micro-ultrasound system.
  • Target Domain (Evaluation): Prospective data from 118 patients collected between 2021 and 2024 across two separate clinical sites:
  • Center B: 76 patients (1,118 cores), 34.6% cancer prevalence.
  • Center C: 42 patients (569 cores), 27.0% cancer prevalence. Both target centers utilized a newer-generation ExactVu micro-ultrasound system and featured different biopsy protocols (systematic vs. targeted).

Key Insights and Ideas

  • Anatomical Invariance as an Anchor: Relying on a structured anatomical prior (the prostate boundary) provides a significantly more stable and targeted adaptation signal than generic self-supervised objectives like entropy minimization.
  • Domain Shift is Encoded Early: Ablation studies revealed that updating the first $n$ encoder layers outperformed updating middle or late layers. This strongly implies that hardware-driven distribution shifts primarily manifest as low-level feature perturbations.
  • Robustness to Imperfect Supervision: Surprisingly, ANT remains highly effective even when the pseudo-masks are imperfect. In cases where the segmentation model achieved a low Dice score (median ~0.60), ANT still consistently improved clinically significant cancer detection over the unadapted baseline.

Key Results

Evaluated under a rigorous leave-one-center-out protocol, ANT achieved top-tier performance:

  • Detection Accuracy: ANT improved the mean Area Under the Curve (AUC) by 2.9% at the biopsy-core level and 3.6% at the patient level compared to the non-adapted baseline, outperforming established TTA methods like TENT, SAR, and CoTTA.
  • Spatial Specificity: ANT drastically reduced "activation leakage"—spurious model predictions outside the actual biopsy needle region—ensuring that the model's focus remained tightly confined to relevant anatomy.
  • Architecture Agnostic: The authors successfully applied ANT to ProstNFound+ (the current state-of-the-art model utilizing a MedSAM backbone), yielding a mean AUC improvement of over 3% without any architecture-specific hyperparameter tuning.

Limitations and Future Work

The primary limitation of ANT is its reliance on an external, pretrained prostate segmentation model to generate pseudo-labels. While the framework proved robust to lower-quality masks, its performance under more extreme domain shifts—such as entirely different scanner manufacturers or highly divergent patient populations—remains unverified. Future work will focus on evaluating ANT against these more severe distribution shifts and exploring its applicability to other organs and modalities.

Summary updated

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.
BibTeX
@misc{dzikunu2026learning,
  author = {Obed Korshie Dzikunu and Mohammad Mahdi Abootorabi and Mohamed Harmanani and Paul F. R. Wilson and Emma Willis and Ferdinand Luger and Adam Kinnaird and Brian Wodlinger and Parvin Mousavi and Purang Abolmaesumi},
  title = {{Learning Prostate Anatomy at Test Time for Cancer Detection in Micro-Ultrasound}},
  year = {2026},
  eprint = {2608.20557},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.20557},
}
Figure