PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ayusha Abbas, Saram Abbas, Kabita Adhikari

arXiv:2608.14796v1cs.CVcs.LG

TL;DR

Accurate prostate boundaries are important for PSAD calculation and targeted biopsy, but 29 MHz micro-ultrasound segmentation remains unevaluated and prompt-sensitive. This paper evaluates a zero-shot MedSAM pipeline with training-free shape refinement, achieving clinician-comparable overlap and 38–52% lower inter-patient variability than human raters.

  • Problem

    29 MHz micro-ultrasound prostate segmentation lacks evaluation, while foundation-model performance varies substantially by modality and prompt quality despite the clinical need for accurate gland boundaries.

  • Method

    The pipeline combines frozen MedSAM inference with CLAHE enhancement and training-free Fourier contour smoothing, evaluating bounding-box and point-click prompting.

  • Results

    38–52% lower inter-patient standard deviation was observed than for human raters, alongside a 45% reduction in HD95 boundary-distance error.

  • Takeaways & Limitations

    Zero-shot foundation-model segmentation can provide a more consistent baseline than variable human annotations for specialized micro-ultrasound prostate analysis.

  • Takeaways & Limitations

    Clinical deployment remains limited by manual bounding-box dependence, and the evaluation cohort uses a single scanner vendor.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper introduces the first systematic, training-free evaluation of MedSAM on 29 MHz micro-ultrasound prostate segmentation, a speckle-dominated modality previously untested with foundation models. It proposes bounding-box prompting and post-hoc geometric filtering to address prompt sensitivity and acoustic speckle without weight optimisation.

  • Clinical motivation: Over 375,000 prostate cancer deaths occur annually, while exact gland boundaries underpin PSAD calculation and targeted biopsy.Conventional TRUS scanners miss roughly one in three high-risk cancers because of poor spatial resolution.
  • Clinical motivation: 29 MHz micro-ultrasound provides three to four times the spatial resolution of 6–12 MHz ultrasound but introduces speckle-related segmentation challenges.The modality matches MRI sensitivity for clinically significant prostate cancer, while conventional ultrasound blurs the prostate capsule into surrounding tissues.
  • Prompt strategy: Bounding-box prompts are necessary for acoustic speckle, whereas point-click prompts collapse irrespective of click number or placement (p<0.00001).The paper attributes this failure to wave-interference speckle rather than dataset-specific artefacts, suggesting transfer to other speckle-dominated modalities and newer prompt-based architectures.
  • Training-free adaptation: 38–52% lower inter-patient variability is achieved by training-free post-hoc Fourier contour smoothing rather than weight optimisation, matching non-expert clinical-rater annotation quality.This reframes domain adaptation as geometric filtering instead of supervised parameter training.

II. RELATED WORK · A. Medical Vision Foundation Models

SAM established zero-shot, prompt-driven segmentation across arbitrary domains, but its medical-imaging performance was initially inconsistent. MedSAM addressed this limitation by fine-tuning SAM’s mask decoder on over 1.5 million image-mask pairs spanning diverse modalities.

  • A. Medical Vision Foundation Models: SAM demonstrated zero-shot prompt-driven segmentation across arbitrary domains, marking a paradigm shift in computer vision.
  • A. Medical Vision Foundation Models: SAM’s medical-imaging performance was initially inconsistent because its pre-training set lacked specialized anatomical features.
  • A. Medical Vision Foundation Models: Over 1.5 million image-mask pairs from CT, MRI, endoscopy, and other modalities were used to fine-tune MedSAM’s mask decoder.

B. Prostate Segmentation on Ultrasound … B. Model Architecture and Zero-Shot Inference

The study evaluates zero-shot prostate segmentation on a 29 MHz micro-ultrasound dataset using a frozen MedSAM architecture with standardized slice preprocessing and bounding-box prompting. The evaluation spans 75 patients and 2,621 valid axial slices, with a 20-patient held-out clinical benchmark.

  • B. Prostate Segmentation on Ultrasound: Historically, prostate ultrasound segmentation has relied on fully supervised architectures, including MicroSegNet’s 3D CNNs for axial-slice continuity.MicroSegNet is described as achieving high segmentation accuracy on micro-ultrasound data.
  • A. Dataset and Patient Cohort: 75 patients and 2,621 valid axial slices comprise the public Micro-Ultrasound Prostate Segmentation dataset acquired at 29 MHz.Volumes contain 21–45 axial slices at 1372×962-pixel resolution.
  • A. Dataset and Patient Cohort: 20 held-out patients provide the clinical benchmark, using four independent annotators.The supplied passage introduces this benchmark but truncates its further annotation details.
  • B. Model Architecture and Zero-Shot Inference: MedSAM uses a ViT-B/16 image encoder, prompt encoder, and lightweight mask decoder, with all pre-trained weights frozen.This defines the zero-shot inference architecture without task-specific training.
  • B. Model Architecture and Zero-Shot Inference: Each axial slice is normalized to [0, 255], converted to RGB, and anisotropically resized to 256×256 pixels.Resizing uses bilinear interpolation.
  • B. Model Architecture and Zero-Shot Inference: The oracle bounding-box prompt expands the tight ground-truth boundary by 20 px in the original 1372×962 resolution space.The prompt is defined before resizing and supplies MedSAM’s spatial input.
  • B. Model Architecture and Zero-Shot Inference: ≈0.8 s/slice is the reported inference time on a standard laptop CPU using Metal Performance Shaders GPU acceleration.The implementation passage specifies MPS acceleration for inference.

C. Training-Free Post-Processing Pipeline

The training-free pipeline uses CLAHE, binary dilation, and Fourier contour smoothing to address speckle, undersegmentation, and jagged boundaries. It retains four low-frequency modes and scales the smoothed contour by s = 1.05 to recover obscured lateral pixels.

  • Contrast Enhancement: CLAHE uses a 0.03 clip limit and 8×8 tile grid to suppress acoustic speckle and sharpen the outer boundary.
  • Morphological Dilation: 3 dilation iterations with a 2D 4-connectivity cross-shaped element recover capsular pixels missed by the raw model.
  • Fourier Contour Smoothing: M = 4 low-frequency Fourier modes are retained to suppress jagged boundary artefacts while preserving characteristic anatomical shape.The largest connected component boundary is represented as a closed curve and treated as a 1D complex spatial signal.
  • Fourier Contour Smoothing: s = 1.05 uniformly scales the reconstructed contour about its centroid to recover lateral boundary pixels obscured by acoustic shadowing.The smoothed contour is reconstructed using the inverse discrete Fourier transform of the truncated spectrum.

D. Automatic Bounding Box Localisation

The section evaluates two automatic bounding-box localisers to reduce manual prompt dependence in clinical deployment. These comprise a statistical centroid-and-size prior and a two-pass MedSAM cascade that tightens the prompt around a coarse prediction.

  • Automatic localisers: Two automatic localisers are evaluated to assess clinical feasibility without relying on manual prompts.The approaches are a statistical centroid-and-size prior and a two-pass cascade model.
  • Statistical Centroid and Size Prior: The statistical prior derives normalized centroid and box dimensions from 55 training patients and applies them to each test slice’s image dimensions.The prior uses average normalized centroid coordinates (µx, µy) and dimensions (µw, µh) relative to image size.
  • Two-Pass Cascade Model: The two-pass cascade first uses the static prior, then prompts MedSAM with a tight box around the coarse mask after adding 20 px padding.The second-pass segmentation proceeds through the final post-processing pipeline.

IV. RESULTS · A. Ablation of Post-Processing Stages · B. Performance and Human Rater Comparison

Post-processing substantially improves MedSAM prostate segmentation, with contrast enhancement and dilation driving overlap gains and Fourier smoothing improving boundary quality. The final pipeline performs strongly across the cohort and held-out test set, with most patients benefiting and no significant mean-Dice difference from human raters.

  • B. Performance and Human Rater Comparison: 0.859 ± 0.037 Dice and 127.0 ± 36.9 px HD95 are achieved across the full 75-patient cohort.The pipeline benefits 74 of 75 patients, with no clinical degradation.
  • B. Performance and Human Rater Comparison: 0.865 ± 0.029 Dice and 120.1 ± 26.1 px HD95 are achieved on the 20-patient held-out test set.These results are reported in the held-out test evaluation.
  • B. Performance and Human Rater Comparison: 29 MHz micro-ultrasound test slices exhibit low signal-to-noise ratio, speckle patterns, and capsule boundary shadowing.Expert prostate masks are shown across 12 representative test patients.
  • B. Performance and Human Rater Comparison: Table II compares held-out-test performance using Holm–Bonferroni-adjusted Wilcoxon tests versus MedSAM, while flagging supervised and oracle-prompt results as non-equivalent or upper-bound references.The table covers N = 20 and identifies the supervised model as not directly comparable and the oracle bounding-box prompt as an upper bound.

C. Prompt Strategy Ablation

The ablation found that simple point prompts performed poorly and evaluated five targeted point configurations. Strategies using background negatives achieved Dice 0.255.

  • Dice = 0.018 for centre-point prompts and Dice = 0.131 for five-point grids, demonstrating poor initial point-prompt performance.
  • Five targeted strategies tested centroid-plus-negatives, dense grids, eroded interiors, random foreground points, and a point with a loose bounding box.The loose bounding box used 100 px padding; grid, eroded-interior, and random strategies each used nine or 25 points as specified.
  • Dice = 0.255 for point strategies with background negatives.

D. Evaluation of Automatic Localisation · V. DISCUSSION

Automatic localisation improves zero-shot prostate segmentation over a statistical prior but remains below the manual-oracle ceiling. The discussion attributes point-prompt failure to speckle, highlights deterministic consistency and clinical relevance, and identifies manual box dependence and scanner-specific evaluation as deployment gaps.

  • D. Evaluation of Automatic Localisation: 0.698±0.058 Dice is achieved by the two-pass cascade, versus 0.691±0.056 for the statistical-prior box and 0.880±0.023 for the manual oracle.The automated strategies remain below the oracle ceiling.
  • V. DISCUSSION: Dice ≤0.35 demonstrates catastrophic point-prompt failure because speckle provides no reliable local structure for individual clicks.The noise-dominated texture field differs from the sharp gradients used by point prompts in CT or MRI.
  • V. DISCUSSION: 0.070 and 0.060 Dice result from dense 9-point and random 25-point prompts, respectively, compared with 0.131 for 5-point prompts.Additional clicks degrade performance by introducing spatial noise, whereas bounding boxes impose an explicit spatial constraint.
  • V. DISCUSSION: Point-prompt sensitivity persists in SAM 2 and LiteMedSAM, whose temporal propagation and memory modules do not remove local coordinate ambiguity.The discussion therefore advises clinical interfaces to avoid point-click inputs for speckle-dominated modalities.
  • V. DISCUSSION: Training-free post-hoc geometric filtering is presented as an alternative to LoRA-based PEFT, whose ultrasound adaptations require hundreds of expert annotations.The proposed adaptation uses Fourier descriptor smoothing rather than parameter updates.
  • V. DISCUSSION: 38–52% lower variability is achieved by deterministic MedSAM inference, with SD = 0.029 versus SD = 0.047–0.061 for human annotators.The discussion links this variability gap to more consistent PSAD calculations and biopsy-needle placement.
  • V. DISCUSSION: The 2025 OPTIMUM randomised trial established 29 MHz micro-ultrasound-guided biopsy as non-inferior to multiparametric MRI-guided biopsy for clinically significant prostate cancer detection.The discussion identifies automated segmentation as relevant for PSAD calculation and biopsy registration as clinical adoption expands.
  • V. DISCUSSION: 0.691 ± 0.056 and 0.698 ± 0.058 from the statistical prior and cascade localisers still leave a gap to the 0.880 ± 0.023 manual-oracle ceiling.Further limitations include a single scanner vendor; proposed future work includes multi-scanner validation and lightweight LoRA fine-tuning on a small labeled dataset.

VI. CONCLUSION

This work presents a zero-shot MedSAM pipeline for 29 MHz micro-ultrasound prostate segmentation that uses frozen inference and training-free Fourier contour smoothing, achieving strong accuracy without backpropagation. Bounding boxes resolve speckle ambiguity where point prompts fail, while the model’s principal clinical advantage is greater consistency than human raters.

  • Zero-shot segmentation: 0.865±0.029 mean Dice was achieved using frozen MedSAM inference with a training-free Fourier contour smoothing prior.The method performs zero-shot segmentation on speckle-dominated 29 MHz micro-ultrasound without modality-specific training.
  • Segmentation performance: 45% reduction in HD95 boundary distance error was obtained without requiring any backpropagation.This result summarizes the pipeline’s boundary-accuracy improvement relative to its baseline.
  • Clinical consistency: 38–52% lower inter-patient standard deviation distinguished the model from non-expert human raters despite no significant difference in average spatial overlap.The conclusion identifies consistency, rather than superior mean accuracy, as the principal clinical advantage.
  • Prompt strategy: Bounding-box prompts are necessary to resolve ultrasound speckle ambiguity, whereas point-click prompts fail because speckle provides no reliable local cues.The prompt ablation found that no point strategy exceeded 0.35, and adding more foreground points reduced performance versus the 5-point baseline.

DECLARATION ON THE USE OF GENERATIVE AI

Generative AI tools assisted with style refinement, grammar checking, and language editing during manuscript preparation. The authors reviewed and edited all generated content and accept responsibility for the manuscript’s scientific integrity and final version.

  • DECLARATION ON THE USE OF GENERATIVE AI: Generative AI tools supported style refinement, grammar checking, and language editing in preparing the manuscript.The authors reviewed and edited all generated content.
  • DECLARATION ON THE USE OF GENERATIVE AI: The authors take full responsibility for the manuscript’s scientific integrity and final version.
Loading 2608.14796v1…