Enhancing Pneumonia Detection from Chest X-Rays Using Probabilistic Pooling with Pretrained CNN Backbones

Abstract

Pneumonia is a significant global health problem, especially among vulnerable populations. Although chest X-rays are commonly used for diagnosis, their interpretation is subjective and prone to variability. Deep learning has improved medical image analysis, while convolutional neural network (CNN) architectures are limited by pooling methods. This work investigates a probabilistic pooling mechanism (ProbPool) that uses softmax-based spatial weighting to include diagnostically salient regions within feature maps. ProbPool is integrated into four widely used pretrained CNN architectures, such as ResNet50, VGG19, EfficientNetB4, and Xception, and evaluated using a consistent training framework on the publicly available Kaggle Chest X-Ray Pneumonia dataset. Experimental results reveal that replacing GAP with ProbPool consistently gives better classification performance across all CNN architectures, giving higher accuracy, AUC, recall, and Dice Similarity Coefficient (DSC).