YOLOv12-WCIRS: An Improved YOLOv12-Based Framework for Small Intestinal Lesion Detection in WCE

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper "YOLOv12-WCIRS: An Improved YOLOv12-Based Framework for Small Intestinal Lesion Detection in WCE" is devoted to the development of a new architecture for detecting small intestinal pathologies in capsule endoscopy images. The authors' comprehensive approach utilizes several modules (WConv, CSFM, IIA_Fusion, RFAHeadDetect, and SD Loss) to simultaneously account for complex backgrounds, low contrast, and small lesion sizes. The paper is well-structured, logical, and provides valuable insights into WCE image analysis. The authors' proposed approach is novel and original. However, it is not without several limitations:

1. The paper provides an overview of modern YOLO approaches, but does not provide a comparison with the latest SOTA or Transformer models. It would be useful to add this to the Introduction.
2. All experiments are limited to the SEE-AI dataset, which limits the quality of validation of the authors' proposed model. The authors note this in the Conclusion, but at least a more detailed discussion of the limitations of the proposed approach is warranted.
3. The discussion section weakly addresses the practical application of the authors' results; appropriate clarifications are needed. This comment is somewhat related to the previous one.

Despite this, the article is a complete work, the content of which is interesting. After addressing the above comments, it can be published in the journal Computers. I recommend accepting the paper after a minor revision.

Author Response

Comment 1:

The paper provides an overview of modern YOLO approaches, but does not provide a comparison with the latest SOTA or Transformer models. It would be useful to add this to the Introduction.

Response 1:

Thank you for this helpful suggestion. We agree that the original Introduction did not sufficiently position the proposed method against the latest SOTA and Transformer-based detectors. In the revised manuscript, we strengthened the Introduction and Related Work sections by expanding the discussion of recent YOLO-series, Transformer-based, and hybrid detection models. We also broadened the experimental benchmark by adding representative Transformer-based and hybrid baselines, including RT-DETRv2 and Deformable DETR, under the same experimental protocol. These revisions improve the background context of the study and better situate the proposed method within the recent object detection literature.

Comment 2:

All experiments are limited to the SEE-AI dataset, which limits the quality of validation of the authors' proposed model. The authors note this in the Conclusion, but at least a more detailed discussion of the limitations of the proposed approach is warranted.

Response 2:

Thank you for this important comment. We agree that relying only on SEE-AI would limit the strength of the validation. In response, we added cross-dataset external validation in the revised manuscript by directly evaluating the model trained on SEE-AI on overlapping lesion categories in Kvasir-Capsule without additional fine-tuning, and the corresponding results are reported in Section 5.2. At the same time, we expanded the Discussion section to more clearly address the limitations of the current study, including that the external validation is restricted to overlapping categories rather than a strict full-class one-to-one benchmark, that the current framework remains a frame-level detector without temporal modeling, and that some rare and visually ambiguous lesions remain challenging. We further clarified that the remaining cross-dataset performance variation is likely related to domain shift arising from differences in capsule devices, illumination conditions, mucosal color representation, and annotation criteria across datasets. We believe these revisions provide a more balanced and transparent discussion of the scope and limitations of the proposed approach.

Comment 3:

The discussion section weakly addresses the practical application of the authors' results; appropriate clarifications are needed. This comment is somewhat related to the previous one.

Response 3:

Thank you for this valuable suggestion. We revised the Discussion section to strengthen the practical interpretation of the results. In particular, we now more clearly explain the clinical relevance of improved lesion sensitivity, reduced false alarms on normal frames, and improved robustness under complex mucosal backgrounds. We also added normal-frame false-alarm analysis and cross-dataset validation results to better reflect practical WCE-assisted reading scenarios, where a large proportion of frames are non-lesion and excessive false alarms may directly increase reading burden. In addition, we now more explicitly discuss the practical trade-off between lesion sensitivity and false alarms, and clarify why reducing false alarms per 10,000 normal frames is clinically meaningful in long WCE examinations. These revisions provide a more complete and realistic account of the practical significance of the reported results.

Reviewer 2 Report

Comments and Suggestions for Authors

The work is promising but requires more justification, validation and analysis before reconsideration.

1. The proposed framework appears to be an incremental integration of existing modules rather than a fundamentally novel architecture. The authors should highlight the motivation behind this. 
2. The paper lacks a clear theoretical justification for combining WConv, CSFM, IIA_Fusion, and RFAHeadDetect into a unified pipeline. 
3. No external validation or cross-dataset benchmarking is performed, which is critical for clinical applicability. 
4. The comparison with state-of-the-art methods is not fully fair, as implementation details and hyperparameter tuning for baselines are insufficiently described. 
5. The improvement over YOLOv12 (6.5% mAP@50) may be attributed to increased architectural complexity rather than principled innovation. 
6. Computational efficiency claims are weakly justified, as FLOPs and FPS comparisons lack normalized hardware benchmarking. 
7. Ablation studies do not sufficiently isolate the contribution of each module under controlled conditions. 
8. The method ignores temporal information in WCE videos, which is a critical aspect for lesion detection improvement. 
9. The paper does not provide model interpretability or explainability analysis, which is crucial for medical AI adoption.
10. The comparison lacks Transformer-based or hybrid architectures, which are strong recent baselines in medical imaging.
11. There is no error analysis (false positives/false negatives breakdown) to understand failure modes.

Author Response

Comment 1:

The proposed framework appears to be an incremental integration of existing modules rather than a fundamentally novel architecture. The authors should highlight the motivation behind this.

Response 1:

Thank you for this important comment. We agree that the original manuscript did not sufficiently clarify the nature of the contribution. In the revised manuscript, we explicitly repositioned YOLOv12-WCIRS as a task-oriented redesign of YOLOv12 for WCE lesion detection rather than a fundamentally new detector architecture. We strengthened the motivation in the Introduction and further clarified in Section 4.4 that the contribution of this work lies in the coordinated integration of complementary modules to address specific WCE challenges, including subtle lesion appearance, weak boundaries, complex mucosal background, and scale imbalance. This revision makes the motivation and contribution of the proposed framework more precise and better aligned with the actual scope of the work.

Comment 2:

The paper lacks a clear theoretical justification for combining WConv, CSFM, IIA_Fusion, and RFAHeadDetect into a unified pipeline.

Response 2:

Thank you for this valuable suggestion. We agree that the original version did not sufficiently explain the rationale for combining these modules into a unified framework. In the revised manuscript, we substantially strengthened the methodological justification in Section 4 by reorganizing the proposed method into a stage-wise pipeline, including backbone local enhancement, neck-based selective fusion and refinement, and receptive-field-aware prediction with scale-aware optimization. We also added an Integrated Discussion subsection in Section 4.4 to explain that the modules were not combined arbitrarily, but were introduced to address sequential bottlenecks in WCE lesion detection, namely insufficient local sensitivity, mismatch between shallow and deep features, background interference, localization instability, and scale imbalance. In particular, we now clarify the sequential mechanism more explicitly: WConv enhances weak and fine-grained lesion responses at the early stage, CSFM preserves the complementarity between shallow structural details and deep semantic cues during cross-scale aggregation, and IIA_Fusion suppresses distracting activations caused by folds, bubbles, residue, and specular reflections so that lesion-relevant details remain distinguishable in complex mucosal backgrounds before RFAHeadDetect and SD Loss further improve prediction and optimization. Although we do not claim a formal theoretical proof, the revised manuscript now provides a much clearer task-driven and mechanistic justification for the unified design.

Comment 3:

No external validation or cross-dataset benchmarking is performed, which is critical for clinical applicability.

Response 3:

We fully agree with this comment. To address this concern, we added a new cross-dataset external validation experiment in Section 5.2, where the model trained on SEE-AI was directly evaluated on overlapping lesion categories in Kvasir-Capsule without additional fine-tuning. The corresponding results are reported in Table 3. In addition, we further included normal-frame false-alarm evaluation on both SEE-AI and Kvasir-Capsule, as shown in Tables 4 and 5, to better reflect practical clinical applicability in WCE-assisted reading. We also explicitly acknowledge in the Discussion section that this external validation is limited to overlapping categories rather than a full-class one-to-one benchmark, and we further discuss that the remaining performance variation is likely related to domain shift caused by differences in capsule devices, illumination characteristics, mucosal color representation, and annotation criteria across datasets.

Comment 4:

The comparison with state-of-the-art methods is not fully fair, as implementation details and hyperparameter tuning for baselines are insufficiently described.

Response 4:

Thank you for pointing this out. We revised the manuscript to improve the transparency and fairness of the comparison protocol. In Section 4.6, we now provide a much more detailed description of the experimental settings, including the examination-level split strategy, input resolution, batch size, learning rate, weight decay, momentum, AMP setting, confidence threshold, IoU threshold, and augmentation strategy. We also clarified that all comparative methods were trained and evaluated under a unified protocol using the same split, preprocessing pipeline, and evaluation criteria. These revisions were made to make the benchmarking process clearer and more reproducible.

Comment 5:

The improvement over YOLOv12 (6.5% mAP@50) may be attributed to increased architectural complexity rather than principled innovation.

Response 5:

We appreciate this concern and revised the manuscript accordingly. First, in the revised comparative experiments, we report not only mAP but also GFLOPs, parameter count, and FPS, so that the performance gain can be assessed together with computational complexity. Second, we strengthened the controlled ablation study in Section 5.4 by including both single-module and multi-module variants under the same protocol. These results show that the improvement is not caused by a single arbitrary addition, but rather by complementary gains from local enhancement, selective fusion, prediction refinement, and scale-aware optimization. In addition, we moderated the novelty claim throughout the manuscript and now describe the method as a coordinated WCE-oriented redesign rather than a fundamentally new architecture.

Comment 6:

Computational efficiency claims are weakly justified, as FLOPs and FPS comparisons lack normalized hardware benchmarking.

Response 6:

Thank you for this comment. We agree that efficiency claims should be carefully presented. In the revised manuscript, we clarified in Section 4.6 that all FPS values were measured on the same RTX 4070 platform under identical inference settings, including the same input size, confidence threshold, IoU threshold, and software environment. We also report Params and GFLOPs together with FPS in Table 2 to provide a more complete efficiency profile. At the same time, we revised the wording in the manuscript to avoid overgeneralized claims and now describe the model as having competitive efficiency under the same hardware setting rather than claiming universally normalized deployment superiority across different platforms.

Comment 7:

Ablation studies do not sufficiently isolate the contribution of each module under controlled conditions.

Response 7:

We appreciate this suggestion and revised the ablation study accordingly. In Section 5.4, all ablation variants are now explicitly described as being trained and evaluated under the same data split, preprocessing pipeline, augmentation strategy, optimization settings, and inference thresholds, with only the module configuration changed. We further expanded the ablation table to include the direct baseline, all single-module variants, and multiple combination variants. This revision allows the contribution of each module to be more clearly isolated and demonstrates that the final improvement arises from cumulative and complementary effects rather than uncontrolled architectural accumulation.

Comment 8:

The method ignores temporal information in WCE videos, which is a critical aspect for lesion detection improvement.

Response 8:

We agree that temporal information is an important direction for WCE analysis. However, the current study focuses on frame-level lesion detection because reliable temporal annotation and sequence-level modeling remain challenging in long WCE videos with intermittent lesion visibility and large frame redundancy. In the revised manuscript, we explicitly acknowledge this limitation in the Discussion section and clarify that the present framework is designed as a spatial detector rather than a temporal-spatial model. We also added a statement in the Discussion and Conclusion that future work will explore temporally aware capsule endoscopy analysis. Guided by the current error analysis, we further note that tiny lesions and visually ambiguous cases remain important sources of missed detections, and future work will therefore explore super-resolution enhancement, rare-class strengthening, and temporally consistent modeling to better capture subtle lesion cues.

Comment 9:

The paper does not provide model interpretability or explainability analysis, which is crucial for medical AI adoption.

Response 9:

Thank you for this helpful suggestion. In response, we added an interpretability analysis in Section 5.3 using Grad-CAM visualization. The corresponding results are shown in Figure 9, where the lesion attention maps of YOLOv12 and YOLOv12-WCIRS are compared. The revised manuscript now discusses that the proposed model produces more concentrated lesion-oriented activation while reducing irrelevant responses in surrounding non-lesion regions. We believe that this addition improves the interpretability and practical relevance of the study for medical AI applications.

Comment 10:

The comparison lacks Transformer-based or hybrid architectures, which are strong recent baselines in medical imaging.

Response 10:

We agree with this comment and expanded the benchmark accordingly. In the revised manuscript, Section 5.1 now includes RT-DETRv2 as a representative real-time Transformer/hybrid detector and Deformable DETR as a representative Transformer-based detector, in addition to the YOLO-family baselines. These results are summarized in Table 2. We also revised the surrounding discussion to explain why broader architectural comparison is important in this task and how the proposed method performs under this expanded benchmark setting.

Comment 11:

There is no error analysis (false positives/false negatives breakdown) to understand failure modes.

Response 11:

We appreciate this comment and agree that failure-mode analysis is important, especially for medical image detection tasks. In the revised manuscript, we added a dedicated error analysis subsection in Section 5.3. Specifically, we now provide a false-negative breakdown in Table 6 and a false-positive breakdown in Table 7, covering representative error categories such as tiny lesions, low-contrast lesions, ambiguous boundaries, mucosal folds, bubbles, residue, and reflections. We also included qualitative examples in Figure 8 to further illustrate these failure modes and the differences between YOLOv12 and YOLOv12-WCIRS. In the revised Discussion, we further explain that these residual errors directly inform future improvement directions, especially for tiny lesions and visually confusing cases. These additions provide a clearer understanding of where the model improves, where challenges still remain, and how the current findings can guide subsequent model refinement.