acceptodds
Under review as a conference paper at ICLR 2027

VCRE-Fib: View-Conditioned Regional Evidence for Fine-Grained Ultrasound Grading of Schistosoma japonicum-Associated Liver Fibrosis

Abstract

Accurate assessment of Schistosoma japonicum-associated liver fibrosis is essential for disease management and long-term follow-up in endemic regions. Ultrasound provides non-invasive imaging, but complex local echogenic patterns and anatomical structures make fine-grained grading challenging. Existing deep learning methods can predict fibrosis scores, yet directly incorporating acquisition views and regional cues into grading while retaining spatial information for inspection remains an open problem. Here we present VCRE-Fib, a view-conditioned regional evidence framework that integrates anatomical context, local information, and global image assessment for fine-grained ultrasound grading. The framework forms view-conditioned local grading evidence before spatial pooling, uses weak localization to guide its aggregation, and combines it with global predictions. Image-only inference jointly returns a fibrosis score, acquisition view, and candidate abnormal-region map. We developed and evaluated the method on a re-curated cohort of 108,709 ultrasound images from 6,373 patients across 35 centers. On a patient-disjoint test set of 4,107 images from 240 patients across four centers, VCRE-Fib reduced the prespecified composite grading risk by 7.115% relative to SFibAI trained and evaluated on the same data split. Image-level mean absolute error decreased from 0.391 to 0.378, alongside lower patient-max, patient-median, and center-balanced risks. The full model also achieved lower composite grading risk than variants that separately removed view conditioning or weak localization. These results support incorporating anatomical context and regional evidence into ultrasound grading while exposing spatial predictions for inspection alongside severity estimates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
medical image analysis, ultrasound, liver fibrosis grading, view-conditioned representation learning, regional evidence, weakly supervised localization, multi-task learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_4s4tOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a evidence framework dubbed VCRE-Fib, for ultrasound grading of a specific subtype of liver fibrosis. The pipeline combined a global predictions and weakly supervised localization. Experiments are conducted on a large dataset comprising 108709 images from 6373 patients. On the test set of 4107 images from 240 patients, the proposed method improves the grading risk.

Critical strengths
  1. The paper addressed a clinically relevant problem using a large dataset for a specific disease.

  2. The paper is easy to follow.

Critical weaknesses
  1. The proposed VCRE-Fib combined existing components, the core innovation is that view conditioning is beneficial. The authors did not establish whether the six conditional branches learn representations at different level. The overall novelty of this paper is compromised.

  2. The full model achieved 66.81% accuracy, compared with 75.70% without weak localization, but its localization Dice decreases from 0.431 to 0.360. This suggests potential negative effect in tasks and grading.

  3. In its current form, the work appears more aligned with the scope of a medical image computing conference such as MICCAI, rather than ICLR.

Questions for authors
  1. As the primary evaluation relies on a risk metric, but the classification performance is less extensively analyzed. Since fibrosis grading is an ordinal classification problem, it is necessary to provide a more comprehensive evaluation using ordinal classification metrics, e.g., such kendall tau, confusion matrices. In particular, it would be important to understand whether the proposed method reduces clinically significant misclassification across severity grades.
Reviewer_aXtNOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VCRE-Fib, a framework for fine-grained ultrasound grading of Schistosoma japonicum-associated liver fibrosis. The method extends the SFibAI grading framework by incorporating acquisition-view prediction and weakly supervised regional localization. A shared ResNet-50 encoder extracts global and spatial features, while six view-specific adapters generate regional grading evidence before spatial aggregation. The resulting regional corrections are combined with global grading logits, and view-conditioned predictions are mixed according to predicted view probabilities.

The method is evaluated on a retrospective multicenter cohort comprising 108,709 ultrasound images from 6,373 patients. On a patient-disjoint test set of 4,107 images from 240 patients, VCRE-Fib reduces the composite grading risk by 7.115% relative to the retrained SFibAI baseline. Additional experiments examine module deletion, patient-level grading, view classification, weak localization, and computational efficiency.

Critical strengths
  1. The paper addresses a clinically meaningful problem using a relatively large multicenter ultrasound dataset. The patient-disjoint evaluation, matched SFibAI baseline, and reporting of image-level, patient-level, and center-balanced grading metrics provide a useful assessment of model performance beyond conventional classification accuracy.

  2. The proposed framework explicitly incorporates anatomical view information and regional evidence into the grading computation, rather than using localization only for post-hoc visualization. The full model achieves consistent improvements over SFibAI across several grading metrics while additionally producing inspectable view and localization outputs.

Critical weaknesses
  1. The contribution of the proposed view-conditioned regional aggregation is not sufficiently isolated. The module-deletion experiments simultaneously change auxiliary supervision, model capacity, and information routing. Therefore, the reported improvements cannot be attributed specifically to view conditioning or localization-guided aggregation. Controlled comparisons with matched-capacity multitask baselines and variants using view or localization supervision without feeding their predictions into grading are needed.

  2. The empirical evidence is insufficient to establish the robustness and practical value of the proposed components. The experiments use a single random seed without statistical significance testing, despite a relatively small composite-risk difference between the full model and its strongest ablation. Furthermore, the full model substantially underperforms the corresponding ablations in view classification accuracy (66.81% versus 75.70%) and localization Dice (0.360 versus 0.431). These trade-offs require further explanation, particularly because reliable spatial predictions are presented as a key benefit of the framework.

Questions for authors
  1. Have the authors evaluated a matched-capacity multitask baseline that predicts fibrosis grade, acquisition view, and localization without incorporating the latter two outputs into the grading pathway? This comparison would help isolate the contribution of the proposed view-conditioned regional aggregation mechanism.

  2. All experiments appear to use a single random seed (2026), without statistical significance testing. Are results from multiple independent training runs available? Reporting mean performance and standard deviations across random seeds, together with patient-level bootstrap confidence intervals or paired significance tests, would help establish whether the improvements over SFibAI and the strongest ablation are statistically reliable.

  3. Why does the full model substantially underperform the corresponding ablations in view classification accuracy (66.81% versus 75.70%) and localization Dice (0.360 versus 0.431)? Could the authors clarify whether these differences arise from multitask optimization interference or limitations of the proposed evidence aggregation?

  4. Has the proposed framework been evaluated with a more recent visual backbone, such as ConvNeXt or Swin Transformer? Such results would help determine whether the gains generalize beyond the ImageNet-pretrained ResNet-50 used in the current experiments.

Reviewer_voNTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Adds a regional-evidence branch to SFibAI (ResNet-50, 36-level 0.0–3.5 ultrasound score for S. japonicum liver fibrosis). Six view-specific adapters produce per-location 36-bin logit corrections before pooling, weighted by a weak-localisation map times learned support, summed and added to the global logits, then mixed by predicted view probabilities; multitask loss adds view cross-entropy and a box-based weak-localisation term. Inference takes one image and returns score, view, and localisation map.

Critical strengths
  1. The evaluation protocol is disciplined: patient-disjoint splits, checkpoints frozen on validation before test, identical seed and recipe for all models, full resource measurements (Table S6), and the authors state plainly that the deletion comparisons cannot separate supervision, capacity and routing.
  2. The task matters for public health, the cohort is large (109k images), and the score + view + localisation interface is a sensible direction for clinical review.
Critical weaknesses
  1. The baseline cannot be trusted, so the claim does not stand. SFibAI’s own paper reports MAE 0.116 with 93.9% within ±0.5; retrained here it gets 0.391 and 69.6%, a 3.4× gap. Appendix A.2 explains that the authors replaced SFibAI’s crop/augmentation (the original uses expert boxes to guide random cropping), initialisation and selection rule; the main text never cites the published numbers or explains the gap. SFibAI also used the boxes, just differently; under the authors’ recipe it no longer does, and is then compared with a method that does. No generic alternative (attention-MIL, FiLM, a plain multitask head) is compared, so the gain cannot be attributed to the design rather than to two extra auxiliary losses.
  2. The effect is small and unsupported. MAE improves by 0.013, from a single seed, with no confidence intervals and, by the authors’ own statement, no significance testing. The −7.1% headline is on a two-layer bespoke composite (R_final) that was also the checkpoint-selection criterion. Patient-level max/median aggregate reference and prediction separately rather than pairing per image. All four test centres appear in training, and one has two patients. The spatial outputs are weak: the full model’s view accuracy is 66.8% and Dice 0.36, versus 75.7% and 0.43 for the deletion variants, so one in three “inspectable” view labels is wrong.
Questions for authors
  1. Please rerun SFibAI with its own released recipe (box-guided cropping and its checkpoint selector) on this split, and explain why the retrained baseline is 3× worse than the published MAE of 0.116.
  2. Please provide multi-seed results with patient-level bootstrap intervals, differences on standard ordinal metrics (MAE, QWK, Spearman), and a generic control (same backbone + attention-MIL head, or SFibAI + view head + box head without the regional branch).
  3. What justifies the weights in R_final and COR, and why is R_final used both for checkpoint selection and as the headline? Given that the full model’s view accuracy and Dice are both below the deletion variants, what is the argument for showing these spatial outputs to clinicians?
Reviewer_1xaUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VCRE-Fib for fine-grained ultrasound grading of Schistosoma japonicum-associated liver fibrosis. On a patient-disjoint test set of 4,107 images from 240 patients across four centers, VCRE-Fib achieves a composite risk of 0.187 compared with 0.201 for SFibAI, corresponding to a 7.115% relative reduction. It also improves image-level MAE from 0.391 to 0.378 and four-grade accuracy from 64.11% to 67.03%. The deletion experiments further show that the full model’s lower grading risk coexists with lower view-recognition and weak-localization scores than the corresponding variants.

Critical strengths

The study utilizes a very large dataset. It uses 108,709 images from 6,373 patients across 35 centers, with patient-disjoint splits.

Critical weaknesses

The main concern is the limited methodological novelty. Although VCRE-Fib introduces a specific integration of view conditioning and localization-guided regional evidence, its main building blocks, including multitask supervision, view-specific adapters, regional weighting, and probability mixing, are established techniques. The novelty therefore lies primarily in their task-specific combination rather than in a new learning mechanism. Moreover, the comparisons with SFibAI and the module-deletion variants do not establish that this combination is preferable to simpler conditioning or aggregation alternatives. The methodological advance is therefore incremental and insufficiently distinguished from existing approaches.

Questions for authors
  1. VCRE-Fib appears to combine several established techniques. What is its core methodological novelty beyond this combination?

  2. The current ablations change supervision, model capacity, and information routing at the same time. Can the authors demonstrate that a simpler multitask or attention-based baseline with comparable capacity cannot achieve the same grading improvement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.