AEC-CapsNet: Enhanced Capsule Networks With Attention and Expansion-Contraction Mechanisms for Advanced Feature Extraction in Medical Imaging

The field of medical image analysis faces challenges due to the complexity of medical data. Convolutional neural networks (CNNs), while popular, often miss critical hierarchical and spatial structures essential for accurate diagnosis. Capsule networks (CapsNets) address some of these issues but struggle with extracting information from complex datasets. We propose the Expanded Attention and Contraction Enhanced Capsule Network with Attention (AEC-CapsNet), designed specifically for medical imaging tasks. AEC-CapsNet leverages attention mechanisms for improved feature representation and expansion-contraction modules for efficient management of global and local features, enabling superior feature extraction. The model is tested on six medical datasets: Jun Cheng Brain MRI (98.14% accuracy, 99.33% AUC), Breast_BreaKHis (98.50% accuracy, 98.96% AUC), HAM10000 (98.43% accuracy, 1.00% AUC), heel X-ray (97.47% accuracy, 99.30% AUC), LC250000 colon cancer histopathology (99.80% accuracy, 99.50%AUC) and LC250000 lung cancer histopathology (99.59% accuracy, 99.20%AUC). Additionally, on the general CIFAR-10 dataset, it achieves 83.48% accuracy, demonstrating robustness and generalisability. To ensure a comprehensive complete assessment, we applied cross-validation for each experiment, which allowed us to evaluate the model's stability and performance across different training datasets. The model was trained for multiple epochs (20, 40, 60, 80, 100, 120 and 140 epochs) to examine its learning and convergence patterns. Without dataset-specific augmentation or architectural modifications, AEC-CapsNet corrects critical weaknesses of existing methods, making it efficient, accurate, and reliable for automated medical image diagnostics.