CNN-BASED DETECTION AND SEGMENTATION OF UTERINE FIBROIDS IN ULTRASOUND IMAGES USING SRAD PREPROCESSING AND GEODESIC ACTIVE CONTOURS

Uterine leiomyomas (fibroids) are the most common benign tumors of the female reproductive system, affecting up to 80% of women by the age of 50 years. Early, accurate, and operator-independent detection remains challenging because of the multiplicative speckle noise inherent in B-mode ultrasound imaging. This study proposes a convolutional neural network (CNN)-based framework that integrates speckle-reducing anisotropic diffusion (SRAD) preprocessing with an attention-enhanced ResNet-50 classifier and geodesic active contour (GAC) segmentation for automated fibroid detection and boundary delineation. A dataset of 1,240 annotated B-mode ultrasound images from 316 patients was used for model development and evaluation. SRAD filtering was applied to suppress speckle noise while preserving tissue boundaries. A modified ResNet-50 incorporating squeeze-and-excitation attention modules was fine-tuned for classification, and Grad-CAM activation maps were used to initialize GAC contour evolution for lesion segmentation. Model performance was evaluated using five-fold stratified cross-validation. The proposed framework achieved an accuracy of 97.4%, sensitivity of 96.1%, specificity of 98.2%, and a Dice similarity coefficient of 95.8%, significantly outperforming three baseline deep-learning architectures (p < 0.01). These results demonstrate that the proposed SRAD–CNN–GAC framework provides a robust and reproducible computer-aided diagnosis approach for automated uterine fibroid detection and segmentation, with strong potential for integration into routine clinical ultrasound practice.