Weakly Supervised Deep Learning for Thoracic Disease Classification and Localization on Chest X-rays

1 INTRODUCTION

• Chest X-ray analysis is challenging because disease appearances vary, while obtaining high-quality lesion annotations requires costly radiologist expertise and data handling.ChestX-ray14 lacks pixel-wise annotations or coarse bounding boxes for most images, so compatible models must rely chiefly on image-level labels and few boxes.Source: §1 · p. 1
• Figure 1 shows radiologist-annotated lesion boxes for eight diseases, and the boxes are used only for evaluation.The figure illustrates the lesion-region annotation that is unavailable for most ChestX-ray14 images.Source: §1 · p. 1, Figure 1 · p. 2
• Prior classifiers using densely connected features do not separate diseases, risking confusion when images contain lesions from multiple diseases.The paper identifies disease-specific lesion areas as a reason to learn disease-discriminative features.Source: §1 · p. 2
• The framework combines DenseNet with squeeze-and-excitation blocks, multi-map transfer, and max-min pooling for weakly supervised disease classification and lesion localization.The authors report superior performance to state-of-the-art methods and verify the proposed components individually on ChestX-ray14.Source: §1 · p. 2
• Squeeze-and-excitation blocks model channel interdependencies, while multi-map transfer learns disease-specific features and max-min pooling aggregates spatial activations.The stated goals are greater sensitivity to subtle normal-versus-lesion differences and improved handling of multi-label cases.Source: §1 · p. 2

2 RELATED WORK

• ChestX-ray14 contains 108,948 frontal images from 32,717 patients, with 14 labels mined from radiology reports rather than manually annotated by pathologists,.Each image may carry multiple disease labels.Source: §2 · p. 2
• ChestX-ray14 labels are noisy; ResNet class-wise ROC-AUC ranged from 0.5609 for Mass to 0.8141 for Cardiomegaly, while high-quality annotations remain scarce.The cited results indicate that performance varied substantially across diseases.Source: §2 · p. 3
• CheXNet, a 121-layer model, outputs disease probabilities and indicative-region heatmaps, but its ImageNet-pretrained representations were not guaranteed to fit thoracic lesions [25].The cited work reported a 0.05 ROC-AUC lift while leaving room for improvement.Source: §2 · p. 3
• A dual-view MIMIC-CXR model concatenated separately trained frontal- and lateral-view features, but lacked a face-to-face comparison with CheXNet on ChestX-ray14 [26].The cited discussion also notes that the numerical results did not strongly support a performance lift from adding views without learning their correlation.Source: §2 · p. 3
• AG-CNN extracts rough regions of interest from a globally trained branch, then combines local-patch and global features for classification [9].The approach addresses small lesions and distortions or misalignment associated with image capture.Source: §2 · p. 3
• A multiple-instance-learning approach treated X-rays as roughly labeled patch bags and used LSE-LBA pooling to create saliency maps at multiple resolutions [35].The cited discussion frames noisy labels and sparse region boxes as a weak-supervision setting.Source: §2 · p. 3

3.1 DenseNet for Chest X-rays

• DenseNet-121 is the backbone because its densely connected layers improve information and gradient flow, and it preserves spatial information.The paper notes that ChestX-ray14 labels are noisy and that images may have multiple disease labels.Source: §3.1 · p. 3

3.2 Squeeze-and-Excitation Block in DenseNet

• The SE block spatially averages convolutional feature maps into a channel descriptor, learns channel-wise attention coefficients, and uses them to recalibrate the features.The coefficients re-scale each channel's transformed feature map.Source: §3.2 · p. 4

3.3 Multi-map Layer and Max-min Pooling

• Multi-map transfer creates M spatially preserved feature maps per disease through 1 × 1 convolution, with image-level labels supervising disease-related modalities.When M = 1, the layer reduces to a standard C-class output.Source: §3.3 · p. 5
• Class-wise average pooling reduces the multi-maps, then max-min pooling combines selected high- and low-activation regions into a global disease score.The pooled score passes through a sigmoid to produce the final positive probability.Source: §3.3 · p. 5

3.4 Comparison with CheXNet

• Compared with CheXNet, the proposed model removes the linear layer, adds SE-based channel recalibration and disease-specific multi-maps, and uses max-min rather than global pooling.The paper presents the fully convolutional design as suitable for spatial learning.Source: §3.4 · p. 5

4.1 Chest X-ray Dataset

• ChestX-ray14 contains 112,120 frontal images across 14 diseases; the patient-level split assigns 76,524, 10,000, and 25,596 images to training, validation, and evaluation.Source: §4.1 · p. 5

4.4 Localization of Lesion Regions

• Radiologist bounding boxes for 880 X-rays were used only for visual localization evaluation, where Figure 4 shows predicted important regions close to corresponding boxes.Source: §4.1 · p. 5, Figure 4 · p. 7

4.2 Experimental Setting

• The model predicts 14 disease probabilities and uses binary cross-entropy; evaluation averages predictions from ten 448 × 448 crops per image.Source: §4.2 · pp. 5–6

4.3 Comparison with State-of-the-art Methods

• Overall performance improved by 2% over previous methods, with at least a 3% margin for challenging diseases such as Lung Nodule.Table 1 reports disease-class and average ROC-AUC scores; Li et al. [22] used extra location information and a nonofficial split.Source: §4.3 · p. 6, Table 1 · p. 6

4.5 Ablation Study

• The full model reached 0.8302 average AUC versus CheXNet’s 0.8180, a reported 1.2% lift from the three components jointly.Source: §4.5 · pp. 6–7, Table 2 · p. 7
• Removing SE blocks lowered average AUC to 0.8279; their squeeze and excitation stages aggregate spatial information and recalibrate convolutional features.Source: §4.5 · p. 7
• With multi-map transfer disabled by setting M = 1, the revised model’s average score was 0.8290.Source: §4.5 · p. 7
• Using only maximum pooling reduced ROC-AUC by 1.0%, supporting the contribution of max-min pooling.Source: §4.5 · p. 7

• The framework jointly classifies thoracic diseases and localizes lesions on chest X-rays using noisy multi-class disease labels.Source: §5 · p. 8
• Its disease-specific multi-map features and cross-channel recalibration produce heat maps that the paper presents as evidence of interpretability.Source: §5 · p. 8
• The paper reports quantitative and qualitative results outperforming state-of-the-art methods, and names more accurate localization using limited bounding boxes as future work.Source: §5 · p. 8