TL;DR
This study investigates whether deeper neural networks provide more reliable visual explanations using Grad-CAM for pneumothorax diagnosis. The researchers found that increasing model depth or switching to Vision Transformers (ViTs) does not necessarily improve diagnostic accuracy or explanation quality; in fact, traditional CNNs often provide more precise and localized heatmaps than their deeper or transformer-based counterparts.
What Problem Does This Paper Address?
Deep learning (DL) has achieved remarkable success in medical imaging, but these models often operate as "black boxes." To build clinical trust, researchers use explainable AI (XAI) tools like Grad-CAM (Gradient-weighted Class Activation Mapping) to produce heatmaps that highlight which parts of an image (like an X-ray) influenced a model's decision.
While Grad-CAM is widely used, there is a lack of systematic research on how different network architectures and depths affect the reliability of these heatmaps. If a model gets the diagnosis right but looks at the wrong part of the lung, it's "right for the wrong reasons"—a dangerous scenario in a clinical setting. This paper benchmarks Grad-CAM across various popular architectures using pneumothorax (collapsed lung) diagnosis as a challenging case study.
Methods
The authors evaluated three major families of DL architectures:
- VGG: (VGG16, VGG19) – Classic "pure" convolutional networks.
- ResNet: (ResNet18, 34, 50, 101) – Networks using skip connections to handle extreme depth.
- Vision Transformers (ViT): (Small, Base, Large) – A newer class of models that use self-attention mechanisms instead of standard convolutions.
To quantify the "goodness" of an explanation, they used two metrics:
- $Diff_{GradCAM}$: The difference between the average heatmap values inside the actual diseased area versus outside. A higher value means the model is focusing more specifically on the pathology.
- Effective Heat Ratio (EHR) AUC: A metric that measures how much of the high-intensity "heat" falls within the ground-truth segmentation of the pneumothorax.
Datasets
The study utilized the SIIM-ACR Pneumothorax Segmentation dataset from Kaggle.
- Total cases: 7,200 (balanced 50/50 between healthy and pneumothorax).
- Ground Truth: Manual segmentations provided by experts for the diseased regions.
- Preprocessing: Images were normalized and enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE).
Key Insights and Ideas
- The Depth Paradox: Deeper is not always better. For ResNets, diagnostic accuracy peaked at ResNet50. Adding more layers (ResNet101) didn't help accuracy and actually changed the "texture" of the visual explanation.
- CNNs vs. Transformers: CNNs (VGG and ResNet) generally produced more localized and accurate heatmaps than ViTs. ViTs tended to produce "dispersed" patterns, likely because their self-attention mechanism looks at global contextual information rather than local pixel gradients.
- The ResNet101 Anomaly: While ResNet101 had high diagnostic accuracy, its Grad-CAM heatmaps often showed two distinct activation areas rather than one focused cluster. This suggests that deeper models might capture more intricate, but potentially less clinically intuitive, representations.
- Visual Mismatch: As shown in Figure 1, even when models agree on a diagnosis, their "reasoning" varies wildly. VGG models often captured the upper chest, while ResNets were more focused on the specific pleural space.
Key Results
- Accuracy: ResNet50 achieved the highest accuracy at 88.20%. ViT models generally performed better than VGG but slightly worse than ResNet.
- Explanation Quality: Despite not having the highest accuracy, ResNet101 and VGG16 led in the EHR AUC metric (0.0319 and 0.0243, respectively), indicating their heatmaps were the most "specific" to the diseased regions.
- ViT Performance: ViT models showed the lowest explanation scores. Interestingly, as ViT models grew in size (from Small to Large), the amount of "irrelevant" area highlighted in the heatmaps increased.
Limitations and Future Work
The authors acknowledge several constraints:
- Dataset Size: The study used a single dataset; ViTs often require much larger datasets to truly shine and might produce better visualizations with more data.
- Natural Image Pre-training: The models were pre-trained on ImageNet (natural images). Fine-tuning on medical data helps, but the "features" learned from natural images might influence the Grad-CAM results.
- Lack of Human Validation: While the quantitative metrics are useful, the authors note that future work should include user studies with clinical experts to see if these heatmaps actually help doctors make better decisions.
- Application Specificity: The findings might differ for other diseases or imaging modalities (like MRI or CT), which remains an area for future exploration.
Summary updated