acceptodds
Under review as a conference paper at ICLR 2027

SGFA: Spatially Guided Feature Adaptation for Multi-Disease Ultrasound

Abstract

Ultrasound lesions from different diseases can have similar intensity and geometry but differ in local texture, making ambiguous boundaries difficult to resolve from spatial cues alone. We propose SGFA (Spatially Guided Feature Adaptation), a multi-disease ultrasound lesion segmentation framework combining Shared Texture Projection (STP) with Quadrant-wise Boundary Box Prompting (QBBP). STP generates image-dependent residual feature updates through projection parameters shared across disease categories. QBBP encodes quadrant-wise boundary boxes as sparse and dense prompts and constrains these updates to prompt-supported regions through a spatial gate. We evaluate SGFA using ground-truth-derived (oracle) prompts on a five-disease pancreatic endoscopic ultrasound (EUS) test set comprising 8,531 frames from 164 patients. We additionally evaluate the framework on the MMOTU ovarian ultrasound dataset with dataset-specific training. SGFA achieves a patient-mean Dice of 0.9548 (frame-mean: 0.9563) on EUS and an image-mean Dice of 0.9642 on MMOTU. Patient-level EUS ablations show gains from STP adaptation under global-box prompting and from jointly adapting STP and the spatial gate under QBBP.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
ultrasound lesion segmentation, multi-disease ultrasound, spatially guided feature adaptation, shared texture projection, boundary box prompting, pancreatic endoscopic ultrasound

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yBzHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SGFA for ultrasound lesion segmentation. It combines Shared Texture Projection (STP) with Quadrant-wise Boundary Box Prompting (QBBP). The prompts are generated from ground-truth lesion contours. The paper evaluates the method on pancreatic and ovarian ultrasound data and reports patient-level results for pancreatic EUS.

Critical strengths

Patient-level evaluation reduces the influence of unequal image counts.

Critical weaknesses
  1. The paper does not justify the main design choices. The reported comparison shows no clear benefit from STP over a simpler adapter. Results for individual diseases are also missing, so the experiments do not establish that the proposed design addresses the disease differences used to motivate it. The need for the main module remains unproven.
  2. The evaluation relies on quadrant boxes built from ground-truth lesion contours, but the paper does not show how these boxes would be obtained when the contour is unknown. The reported results therefore do not establish how the method would perform in practice. The proposed method also receives more ground-truth information than several comparison methods, making it difficult to separate better design from extra information.

These make the claimed advantages insufficiently supported.

Questions for authors
  1. Your pipeline builds quadrant boxes from ground-truth lesion contours, so the segmentation is already known when these prompts are generated. How would this method be used in practice, and how would the boxes be obtained when the contour is unknown?
  2. Do the reported gains remain when all comparison methods receive the same ground-truth information?
  3. Can you provide code and report the numbers of patients and frames in the EUS training and validation sets, along with how the patients were split to train the source checkpoints? These details are missing from the paper and supplement.
Reviewer_kXF5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SGFA combines Shared Texture Projection (STP), a local convolutional bottleneck residual, with Quadrant-wise Boundary Box Prompting (QBBP). Ground-truth-derived quadrant boxes supply sparse and dense prompts and restrict feature updates spatially. The method reports patient-mean Dice of 0.9548 on five-disease pancreatic EUS and image-mean Dice of 0.9642 on MMOTU. Despite high oracle-prompted accuracy, the evidence does not establish a sufficiently novel texture-adaptation mechanism or a substantial benefit beyond richer spatial guidance.

Critical strengths
  1. Relevant evaluation and appropriate patient aggregation. The EUS test cohort includes 164 development-isolated patients across five diagnoses. Equal-patient aggregation and paired patient bootstrapping address unequal frame counts.
  2. Useful transparency and controls. The authors explicitly disclose oracle prompting and provide three-seed results, boundary metrics, component ablations, prompt perturbations, an adapter control, and both improvement and failure examples.
Critical weaknesses
  1. Limited novelty and no demonstrated advantage over a generic adapter. STP uses standard depthwise convolutions, fusion, a low-rank bottleneck, and bounded residuals. Shared weights across diseases are not a distinctive learning mechanism. Supplementary Table 10 reports virtually identical MMOTU Dice for STP and the pointwise adapter (0.964159 versus 0.964157); all four paired metric confidence intervals include zero. The texture-specific design has no demonstrated advantage.
  2. Oracle contour information limits practical relevance. The evaluated QBBP boxes are constructed from ground-truth contour points, providing much richer information than a coarse user box. No clinician-generated or automatically predicted prompts are evaluated. The perturbation study shows sensitivity to missing or inaccurate boxes, but tests only one QBBP-only checkpoint. Robustness of the complete SGFA model under practical prompts remains unestablished.
  3. The incremental adaptation gain is very small. Relative to QBBP only, joint STP-and-gate adaptation improves MMOTU Dice by 0.0583 percentage points and paired EUS patient Dice by 0.3175 points. The confidence intervals support positive effects, but their magnitude is modest compared with the prompting contribution. Together with the adapter control, this weakens the case for a substantial methodological advance.
  4. Existing ablations do not isolate the proposed coupling. STP and the spatial gate are adapted jointly under QBBP; STP training with a fixed gate under the same interface is not tested. Sparse corner tokens, dense occupancy maps, edge maps, and writeback gating are also coupled without separate path ablations. The necessity of the individual mechanisms and their combination remains unclear.
  5. The texture and multi-disease claims lack direct support. Supplementary Table 10(c) finds weak associations between gains and texture-related descriptors. No disease-resolved quantitative comparisons show which diagnoses benefit from STP. Pancreatic cancer accounts for 116 of 164 EUS test patients, while some categories contain only six or nine. Pooled averages cannot establish consistent gains across diseases or a texture-specific mechanism.
  6. The central EUS training protocol is incompletely reproducible. Training/validation patient counts and complete recipes for the global-box and QBBP source checkpoints are not specified for the private EUS cohort. These details matter because adaptation freezes the source encoder, decoder, and prior pathway. The MMOTU recipe does not resolve how to reproduce or assess source-training effects in the main experiment. Also the authors trained all the comparison models by themselves or reported from their original papers and all the mean std values as well?
Questions for authors

See Weaknesses

Reviewer_ZAW5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper builds and trains an ultrasound lesion segmentation system designed to handle a variety of lesion types. They build a complex new architecture, incorporating 2 key modules of theirs: Shared Texture Projection (STP) and Quadrant-wise Boundary Box Prompting (QBBP). However, the validation is on prostate lesions and Ovarian cancer lesions (limiting variability), and QBBP requires human-expert-provided ground-truth bounding boxes as input during inference.

Critical strengths

The authors are pursuing a good goal, and they demonstrate that the combination of their methods and additional human inputs meaningfully improve segmentation accuracy of multiple types of prostate lesions and ovarian cancer lesions. They include numerous performance metrics, which I really appreciate.

Critical weaknesses

As seen in Table 2, QBBP seems be providing almost all the performance gains, whereas STP contributes very little. Furthermore, "QBBP is a GT-derived prompted-segmentation protocol" that requires clinician input to operate, so we should expect a-priori that QBBP would perform better than purely image-based methods. I don't see any fair comparisons against other methods that also rely on human prompting. I also expect to see more diversity in lesion types for a paper with these claims. This paper introduces tremendous complexity and fails to show meaningful gains that are independent of additional clinician inputs.

Questions for authors

Did you compare your performance against any other clinician-prompt-guided segmentation methods, either against ultrasound-specific methods or else against adapted prompt-guided segmentation methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.