Deep Radiomics-Attention Fusion Network (RaAFN) for Accurate Uterine Tumour Classification

Abstract

The complicated structure of uterine tumours and the natural weaknesses of the manual approach to image interpretation result in early-stage classification, which can be challenging to achieve, leading to a delayed diagnosis and poor clinical care. To overcome this issue, the given Radiomics-Attention Fusion Network (RaAFN) incorporates handcrafted radiomics features together with deep learning representations with channel and spatial attention mechanism. Radiomics brings out clinically significant texture, shape, and intensity features, whereas a CNN backbone brings out hierarchical representations of images. The attention modules are used to make the network focus on tumour-relevant regions and a fusion layer is used to combine radiomics and deep features into a single representation that is used to classify uterine tumours robustly using a dense neural network. The provision of this multi-level feature fusion is quite effective to manage image variability and noise, and, therefore, the model can be deployed in real-time clinical use. The experimental assessment is done on a dataset of uterine imaging; it tends to classify better with 94% accuracy surpassing the available methods. RaAFN is the most important innovation in the sense that it takes into consideration the attention-directed integration of the hybrid features and it offers an accurate, interpretable and clinically reliable framework of automated uterine tumour classification that facilitates timely diagnosis and the post-processing of improved patient outcomes.

Keywords
Attention Mechanism, Deep Learning, Feature Fusion, Radiomics, Uterine Lesion Classification, Uterine Tumour Diagnosis