CBAM-Enhanced DenseNet121 for Multi-Class Chest X-Ray Classification with Grad-CAM Explainability

A new research paper proposes CBAM-DenseNet121, an AI model designed to tackle a critical shortcoming in automated pneumonia diagnosis. Most existing deep learning systems treat pneumonia detection as a simple binary (yes/no) problem. This new framework, created by researcher Utsho Kumar Dey, goes further by classifying chest X-rays into three categories: Normal, Bacterial Pneumonia, and Viral Pneumonia—a vital distinction for determining the correct antibiotic treatment. The model integrates a Convolutional Block Attention Module (CBAM) into the DenseNet121 architecture, enhancing its focus on relevant image regions. In rigorous testing repeated three times for statistical reliability, it achieved a mean test accuracy of 84.29% with high AUC scores, notably outperforming a baseline study that found even popular models like EfficientNetB3 underperformed in this medical task.

Beyond accuracy, the research emphasizes clinical trust and deployment practicality. The model employs Grad-CAM (Gradient-weighted Class Activation Mapping) to generate visual heatmaps that highlight the areas of the lung the AI used to make its decision. This explainability is crucial for gaining doctor buy-in, especially in low-resource settings like Bangladesh, where the study notes a severe shortage of radiologists. By providing both a precise multi-class diagnosis and a visual rationale, CBAM-DenseNet121 moves beyond a black-box prediction to an interpretable aid that can be integrated into constrained clinical workflows, potentially improving patient outcomes where expert human analysis is unavailable.