Interpretable Chest X-Ray Diagnosis with Deep Learning and Explainable AI

Pneumonia is regarded as a primary cause of death, accounting for 14% of all fatalities in children under five in 2019, representing 740,180 recorded deaths. X-ray imaging provides a low cost and widely available option for pneumonia detection compared to other alternatives and imaging modalities. Automating the diagnosis and interpretability pipeline is imperative in view of the high mortality rates and insufficient number of trained specialists to interpret chest X-rays (CXR). This chapter investigates the techniques for reliably detecting pneumonia from CXR while maintaining interpretability, which is critical for reliable clinical decisions. We used transfer learning with the three cutting-edge Deep Neural Networks (DNN) models, ResNet18, ResNet50, and Inceptionv3 for CXR pneumonia diagnosis and Gradient-weighted Class Activation Mapping (Grad-CAM) to identify the features used by the model for pneumonia identification. These pre-trained models are first fine-tuned on the CXR dataset and then used to design a weighted voting ensemble model that improves the classification accuracy and interpretability. The results show that the ensemble model performed better, with an exceptional accuracy of 97.79% and a precision of 98.02%. The comparative analysis with other models, such as VGG16, VGG19, MobileNetv2, and NasNet, demonstrates the efficacy and robustness of the proposed approach in optimizing pneumonia detection while maintaining interpretability. This chapter demonstrates that combining an ensemble of three high-performing DNN model with Grad-CAM can help clinicians combine their own diagnosis, improving clinical practice and patient care.