An Interpretable Hybrid Deep–Swarm–Capsule Learning Framework For Robust Endometriosis Detection From Ultrasound Imaging

Abstract

Endometriosis is a complicated gynaecological condition, the early diagnosis of which is difficult because of the morphology of lesions that appear to be subtle, noise artefacts in ultrasound, and a low interpretability of traditional deep learning models. In order to overcome these shortcomings, the proposed research suggests a hybrid diagnostic model, combining signal enhancement, swarm-intelligence-based feature optimisation, and capsule-based deep residual learning to make ultrasound-based endometriosis detection more robust. Firstly, it aims at improving the diagnostic relevance of anatomical structures through adaptive Butterworth–wavelet preprocessing and localised segmentation of the those to minimise noises in the shape of speckles whilst maintaining the edges of the lesions. The second goal is to minimise deep feature redundancy and enhance generalisation with the help of an Improved Recursive Bee Colony (IRBC) optimisation algorithm to perform hierarchical feature selection and dimensional refinement. The third goal is to design a Hyper Capsule ResNet50-CNN classifier wherein spatial hierarchies, orientation relationships and contextual dependencies are maintained even when residual learning and capsule routing are involved, as opposed to when using conventional convolutional pooling operations. Rigorous experimental testing on benchmark ultrasound data sets proves that the proposed hybrid model is better than standalone CNNs, ResNets, SVMs, and traditional capsule network models with regards to various performance measures, such as accuracy, sensitivity, specificity, F1-score, and RMSE. These results show that multi-level hybridisation of signal processing, swarm intelligence and deep capsule learning is much more likely to boost diagnostic reliability, robustness, and interpretability, which makes the framework a clinically viable decision-support tool capable of screening early endometriosis without invasive procedures.