A Multi-scale CBAM-ResNeXt Framework for Stage-wise Endometriosis Detection in Ultrasound Imaging

A chronic gynaecological condition that affects a significant proportion of women of reproductive age and often leads to pelvic discomfort, infertility and delayed clinical diagnosis is endometriosis.Accurate identification of disease severity is essential for treatment planning; however, stage-wise assessment using transvaginal ultrasound (TVUS) is challenging due to imaging noise, subtle lesion characteristics and dependence on expert interpretation.This research presents an intelligent deep learning (DL)-based method designed to automatically classify the stages of endometriosis from ultrasound images.A curated dataset of 500 clinically verified TVUS images was collected and categorized into five rASRM-based classes: No Pathology, Stage I (Minimal), Stage II (Mild), Stage III (Moderate) and Stage IV (Severe).To improve image quality, a structured preprocessing pipeline combining hybrid deep denoising, guided filtering and intensity normalization was applied.Dataset imbalance was addressed using a Conditional Deep Generative Adversarial Network (CDGAN), which generated stage-specific synthetic ultrasound images and expanded the dataset to a balanced distribution of 800 images per class.For classification, a hybrid multi-scale CBAM-ResNeXt architecture was developed.The ResNeXt-50 backbone performs hierarchical feature extraction, while a multi-scale convolution module captures lesion patterns at different spatial resolutions using 3×3, 5×5 and 7×7 kernels.A Convolutional Block Attention Module (CBAM) further enhances feature maps by emphasizing diagnostically relevant channels and spatial regions.Experimental assessment validates that the developed framework attains 99.2% accuracy, 99.2% recall, 99.2% F1-score and 99.3% precision.Class-wise performance remained consistently high, with F1-scores fluctuating from 0.988 (Stage I) to 0.996 (Stage IV).Comparative experiments with baseline models confirm the effectiveness of the suggested architecture.These results indicate that the hybrid multi-scale CBAM-ResNeXt approach provides a reliable and automated solution for stage-wise endometriosis classification from ultrasound images.