Explainable and Automated Pneumonia Detection from Chest X-Rays using CNNs

Abstract

Chest X-ray is the most popular examination type for thoracic diseases, but its interpretation exhibits error rates which are still subject to inter-observer variability and economic workload constraints. This work can be found in this paper: "A reproducible deep learning pipeline for pneumonia identification vs normal based on DenseNet-121 from chest X-ray". The dataset originated from the NIH ChestX-ray14 corpus and was downsampled to 8,500 frontal radiographs (1,050 pneumonia-positive, 7,450 normal) and split at the patient level into training, validation and testing sets. Preprocessing: Grayscale normalization, Resize, Targeted Augmentation and Training (with) Early Stopping, Learning Rate Scheduling, Class Weighting and Post-Hoc Probability Calibration. In the held-out test set, the model achieved ROC-AUC: 0.87, PR-AUC: 0.72, as well as a general accuracy of 93.2%, sensitivity: 82.8% and specificity: 94.6%. Calibration analysis contributed to improving the Brier score from 0.042 to 0.019 and led to good-fitting reliability curves. Interpretability was built into the inference using Grad-CAM and Integrated Gradients, with explanation faithfulness quantitatively checked (deletion AUC = 0.84, insertion AUC = 0.87, sanity check pass rate = 98%, pointing-game hit rate = 76%). Based on the above results, it can be seen that CNN-based diagnosis is promising to achieve a good accuracy as well as interpretability and reproducibility simultaneously. Hence, the proposed framework provides a white-box baseline for clinical examination and future multi-label thoracic disease detection extensions