Adaptive Heatmaps for Medical Imaging in Convolutional Neural Networks

We introduce a novel heatmap-based pooling mechanism for convolutional neural networks (CNNs), specifically designed to enhance interpretability in medical imaging tasks. By integrating an adaptive attention mechanism regulated through temperature scaling and entropy-based smoothing, our model dynamically emphasizes diagnostically critical regions while suppressing noise. This approach bridges the gap between high-performance CNNs and the clinical need for transparent AI systems, enabling clinicians to explore fine-grained and coarse-grained regions with greater confidence. The proposed framework is computationally efficient, compatible with existing architectures, and demonstrates significant potential for improving diagnostic accuracy and clinical adoption, offering a robust, scalable, and interpretable solution for high-stakes diagnostic examinations.