Improving Imbalanced Multi-Label Chest X-Ray Diagnosis via CBAM-Enhanced CNN Backbones

Technical Summary: Improving Imbalanced Multi-Label Chest X-Ray Diagnosis via CBAM-Enhanced CNN Backbones

Problem Statement
Chest radiography is a primary modality for diagnosing thoracic diseases, yet manual interpretation remains time-consuming and reliant on expert knowledge. While Deep Learning (DL) has automated feature extraction, two critical challenges persist in this domain:

1. Multi-Label Complexity: A single radiograph often contains multiple co-existing pathologies, requiring models to handle complex, simultaneous classifications.
2. Severe Class Imbalance: Large-scale datasets like ChestXray14, MIMIC-CXR, and CheXpert are dominated by frequent conditions and the "No-Finding" label. This imbalance biases learning behavior, leading to degraded performance on rare but clinically significant diseases.
   Furthermore, standard Convolutional Neural Networks (CNNs) suffer from limited receptive fields, hindering the capture of long-range dependencies crucial for identifying spatially distant but semantically correlated abnormalities. While Transformer-based architectures offer solutions, their high computational and data requirements often limit feasibility in real-world clinical scenarios.

Methodology
The authors propose a practical integration strategy that enhances standard CNN backbones with the Convolutional Block Attention Module (CBAM) to improve multi-label classification without requiring pixel-level annotations. The methodology focuses on three core components:

1. CBAM-Enhanced Architecture:

   • The framework integrates CBAM into widely adopted backbones, specifically DenseNet121 and VGG16.
   • CBAM is applied sequentially along channel and spatial dimensions to generate attention maps that act as soft feature masks. This amplifies task-relevant regions while suppressing irrelevant background information.
   • Strategic Placement: The paper introduces a principled module placement scheme. For DenseNet121, CBAM is inserted after the 3rd and 4th dense blocks. For VGG16, it is inserted after blocks 3, 4, and 5. This strategy targets deeper semantic layers where high-level abstractions are formed, balancing low-level feature extraction with high-level semantic refinement.

2. Training Paradigm:

   • The study adopts a two-stage learning paradigm (inspired by prior work [7]).
   • Stage 1: The model is trained exclusively on abnormal samples to capture class-specific discriminative patterns, mitigating the dominance of the "No-Finding" class.
   • Stage 2: The model is fine-tuned on the entire dataset to reinforce global representation learning.
   • The authors investigate various loss functions, including Binary Cross Entropy (BCE) and Focal Loss, comparing single-stage versus two-stage training.

3. Ablation and Evaluation:

   • The study conducts ablation studies to determine the optimal placement of CBAM modules and the efficacy of different training configurations.
   • Evaluation is performed on the ChestXray14 dataset (112,120 images, 14 disease categories) using a patient-wise split to prevent information leakage.

Key Contributions
The paper outlines four primary contributions:

• Framework Integration: A CBAM-Enhanced CNN framework that incorporates attention mechanisms into standard backbones (DenseNet121, VGG16) to refine feature representations for multi-label classification.
• Placement Strategy: A principled scheme for identifying optimal insertion points for CBAM, demonstrating that integrating attention in deeper semantic layers yields superior results compared to full or minimal placement.
• Training Analysis: An ablation study investigating the effects of Focal Loss, one-stage training, and a curriculum-inspired two-stage strategy. The analysis highlights how training configurations affect performance under severe class imbalance.
• Performance Benchmarking: Evaluation on ChestXray14 showing improved detection of underrepresented conditions compared to state-of-the-art baselines.

Results
The proposed method achieves the following results on the ChestXray14 dataset:

• Overall Performance: The CBAM-Enhanced VGG16 model achieved a mean AUC of 0.8695, and the CBAM-Enhanced DenseNet121 achieved 0.8681. These scores surpass previous benchmarks, including Wang et al. (0.7451), CheXNet (0.841), and SynthEnsemble (0.854).
• Rare Pathology Detection: The models demonstrated resilience to class imbalance in rare categories:
  • Hernia: CBAM-DenseNet121 achieved an AUC of 0.9660 (vs. 0.9172 in prior work).
  • Emphysema: Achieved AUCs of 0.9426 (DenseNet) and 0.9422 (VGG16).
  • Pneumonia: CBAM-VGG16 achieved an AUC of 0.7915.
• Ablation Findings:
  • Placement: Inserting CBAM after the 3rd and 4th blocks in DenseNet121 and blocks 3, 4, and 5 in VGG16 yielded the highest performance. Restricting attention to only the final block caused a significant performance drop.
  • Training Strategy: The two-stage training approach consistently outperformed single-stage training, even when single-stage models utilized Focal Loss. For instance, the two-stage DenseNet121 (BCE Loss) reached 0.8681, significantly higher than the best single-stage result of 0.8400 (Focal Loss). This suggests that isolating abnormal data in the initial stage is critical for learning inter-class semantics in imbalanced datasets.

Significance and Claims
The paper claims that its proposed architecture offers a practical trade-off between modeling capacity and efficiency. By enhancing standard CNNs with attention mechanisms, the method improves the detection of subtle and low-prevalence pathologies without the computational overhead of Transformer-based models or the need for expensive pixel-level annotations. The authors emphasize that the two-stage training curriculum is a critical factor in overcoming class imbalance, allowing the model to effectively learn discriminative patterns before generalizing to the full dataset. The work concludes that targeted deployment of attention mechanisms, combined with tailored training curricula, is essential for robust multi-label diagnosis in medical imaging. Future work is suggested to focus on enhancing clinical interpretability and modeling inter-pathology dependencies.