acceptodds
Withdrawn from ICLR 2027

Unleashing Feature Learning Potential via Tree-like Hierarchical Semantics for Pneumonia Lesion Data Mining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pneumonia lesion data mining inevitably deals with the low-resolution Chest X-Ray (CXR) images, posing a pivotal challenge to effective feature learning. Existing methods resort on complicated network structures to learn the latent features, while suffer from insufficient semantics due to the low resolution, resulting in performance limitations. To address this issue, this paper presents a Tree-like Hierarchical Semantic Enhancement Network (THSE-Net), aiming to unleash the feature learning potential, with an explicit Hierarchical Binary Tree (HBT) and corresponding loss functions. Specifically, THSE-Net innovatively employs the HBT to capture hierarchical semantics, and dynamically regulates the global and local feature maps to align with the semantics; and then, it dynamically computes the HBT semantic attentions based on the enhanced features to improve the feature learning capability. On the publicly available benchmark datasets, the proposed THSE-Net enhanced the IoU metric by 0.79%-1.46%, the DSC metric by 0.9%-1.39%, and the F1 score metric by 0.66-1.15% compared with the existing methods. The code will be released on GitHub upon acceptance.

Primary area:
optimization
Keywords:
feature learning, data mining, binary trees, hierarchical structure, loss functions

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_EhudOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes THSE-Net for lung and infection segmentation and pneumonia prediction from chest X-rays. Its central idea is to encode the containment relation between lung regions and infected regions as a hierarchical binary tree, then use tree-based loss corrections, a hierarchical spatial fusion module, and channel attention during feature learning.

Critical strengths

The paper addresses a failure mode in pneumonia lesion segmentation by incorporating the anatomical relation between lung and infection regions into training. Its comparison with focal tree-min loss and component ablations provide some evidence that the task-specific formulation is useful.

Critical weaknesses

Lack of technical novelty.

The anatomical hierarchy motivating the method is sensible, but the central idea of enforcing parent-child consistency in segmentation has been explored in prior work, including hierarchical segmentation with tree-min loss. And the Proposed Modules (HSFM and CAM) are not that novel.(It's already establish method). Although the paper reports improvements over focal tree-min loss, it does not clearly explain which aspect of the proposed HBT formulation accounts for those gains or establish a sufficiently distinct methodological contribution. This limits my assessment of its significance for ICLR.

Questions for authors

Could the authors clarify the precise mathematical difference between the proposed HBT loss and the focal tree-min loss of Li et al. (CVPR 2022) [1] ?

[1] Deep Hierarchical Semantic Segmentation (Li et al. CVPR 2022)

Reviewer_Kky6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the task of pneumonia lesion mining on chest X-ray (CXR) images. To address the clinical challenge that low-resolution CXR images exhibit insufficient feature learning and are prone to produce false lesion predictions outside lung regions, a tree-like hierarchical semantics enhanced network, THSE-Net, is proposed. A hierarchical binary tree (HBT) is introduced to incorporate medical prior hierarchical semantic constraints, thereby improving pneumonia lesion segmentation and pneumonia classification.

Critical strengths

In this paper, the two-stage clinical workflow of pneumonia diagnosis, namely distinguishing lung from non-lung regions first and then differentiating infected from non-infected regions within the lung, is formulated as a hierarchical binary tree (HBT). Anatomical prior knowledge is explicitly incorporated into network training. Based on the parent--child and sibling relationships among binary tree nodes, three semantic constraint rules are devised, together with a dedicated HBT loss for dynamic adjustment of loss values. By imposing semantic constraints at the loss level, logical control over model outputs is achieved, and the probability of erroneous predictions in which lesions appear in non-lung regions is reduced.

Critical weaknesses

(1) The three handcrafted HBT rules are acknowledged to be necessary rather than sufficient, but typical error cases that these rules cannot constrain or correct are not illustrated, and the resulting degradation of the performance upper bound is not analyzed theoretically.

(2) Hierarchical tree-structured losses have been widely studied in image segmentation, yet only basic losses such as BCE and Focal loss are compared. The differences and novelty of the HBT loss relative to existing hierarchical tree-based losses are not clarified, and comparative experiments are missing.

(3) The HBT loss is formulated as a multi-branch conditional system, but loss convergence and gradient properties are not analyzed. It remains unclear whether gradients are stable across branches during backpropagation, whether gradient vanishing or explosion may occur, and whether the optimization objective is guaranteed to converge.

(4) Most baselines are pneumonia-specific models, and comparisons with general large segmentation models such as SAM are absent. Perturbation tests under noise, image degradation, and CXR images from different devices are also missing, so the stability of the model on noisy and complex real-hospital CXR data is not fully verified.

(5) Model complexity and computational overhead are not evaluated, as parameter counts, inference speed, and FLOPs are not reported. The additional costs introduced by the multi-scale dilated convolutions in HSFM and the 1D dynamic convolutions in CAM are not analyzed, making real-time clinical deployment difficult to assess.

Questions for authors

(1) The three handcrafted HBT rules are acknowledged to be necessary rather than sufficient, but typical error cases that these rules cannot constrain or correct are not illustrated, and the resulting degradation of the performance upper bound is not analyzed theoretically.

(2) Hierarchical tree-structured losses have been widely studied in image segmentation, yet only basic losses such as BCE and Focal loss are compared. The differences and novelty of the HBT loss relative to existing hierarchical tree-based losses are not clarified, and comparative experiments are missing.

(3) The HBT loss is formulated as a multi-branch conditional system, but loss convergence and gradient properties are not analyzed. It remains unclear whether gradients are stable across branches during backpropagation, whether gradient vanishing or explosion may occur, and whether the optimization objective is guaranteed to converge.

(4) Most baselines are pneumonia-specific models, and comparisons with general large segmentation models such as SAM are absent. Perturbation tests under noise, image degradation, and CXR images from different devices are also missing, so the stability of the model on noisy and complex real-hospital CXR data is not fully verified.

(5) Model complexity and computational overhead are not evaluated, as parameter counts, inference speed, and FLOPs are not reported. The additional costs introduced by the multi-scale dilated convolutions in HSFM and the 1D dynamic convolutions in CAM are not analyzed, making real-time clinical deployment difficult to assess.

(6) The CT-based baselines (Emb-trattunet and GAHFNet) are adapted for CXR tasks, but the detailed modifications made for this adaptation are not clearly described.

Reviewer_vhnHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes THSE-Net, a hierarchical semantic enhancement framework for pneumonia lesion analysis using chest X-ray images. The method constructs a hierarchical binary tree representing relationships among lung regions, non-lung regions, infected areas, and non-infected areas. Based on this hierarchy, the framework introduces corresponding loss functions and combines hierarchical spatial feature fusion with channel attention to improve feature representations. Experiments on public pneumonia-related datasets report improvements in lesion segmentation and prediction performance compared with several existing approaches.

Critical strengths
  1. The paper considers an important medical imaging problem. Accurate localization of pneumonia lesions is challenging, particularly when image resolution is limited and lesion boundaries are ambiguous.

  2. Incorporating anatomical relationships into the learning process is a reasonable direction. The proposed hierarchical representation provides an explicit way to distinguish lung regions from non-lung regions and may help reduce anatomically inconsistent predictions.

  3. The authors provide experimental comparisons and ablation studies to evaluate the proposed framework. The reported results suggest that the hierarchical constraints and feature enhancement components can improve segmentation performance.

Critical weaknesses
  1. The methodological novelty is insufficient for a major machine learning conference. The proposed hierarchical binary tree mainly represents predefined anatomical relationships, while the remaining components rely on conventional feature fusion, attention mechanisms, and auxiliary losses. The paper does not clearly establish a substantial methodological contribution beyond existing anatomy-aware segmentation approaches.

  2. The experimental improvements are relatively modest, and the evaluation does not sufficiently demonstrate the necessity of the proposed hierarchical formulation. In particular, it remains unclear whether comparable improvements could be obtained using simpler anatomical constraints, auxiliary segmentation losses, or standard multi-task learning strategies.

  3. The generalization and robustness of the method are not adequately established. The evaluation focuses on a limited set of pneumonia-related benchmarks, without sufficiently investigating performance under different imaging conditions, annotation uncertainty, or external clinical distributions. This limits the strength of the conclusions regarding practical applicability.

Questions for authors
  1. What is the main technical advantage of the proposed hierarchical binary tree over conventional hierarchical segmentation or anatomy-aware loss functions?

  2. Could the authors compare THSE-Net against simpler baselines that enforce lung-region consistency without introducing the complete hierarchical architecture?

  3. How robust is the method to inaccurate lung masks or ambiguous lesion boundaries?

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

Related papers

Discussion (0)

Sign in to comment.