Detecting the Regions-of-Interest that Enclose the Tumors in Breast Ultrasound Images Using the RetinaNet Model
Mohammad I. Daoud, Aamer Al-Ali, Mostafa Z. Ali, Ismail Omar Hababeh, Rami Alazrai · 2023
Ultrasound imaging provides an effective modality to diagnose breast cancer, but the task of interpreting breast ultrasound images is complex and challenging. Computer-aided diagnosis systems can enhance the accuracy of ultrasound-based breast cancer diagnosis, but the creation of these systems necessitates the utilization of effective techniques for detecting the region-of-interest (ROI) that encloses the tumor. The goal of our study is to investigate the effectiveness of utilizing the RetinaNet deep learning model for detecting the ROIs in breast ultrasound images. To the best of our knowledge, the RetinaNet model has not been employed previously to detect the ROIs in breast ultrasound images. To achieve this goal, the ultrasound image is processed using contrast enhancement and sharpening to synthesize an RGB image with the aim of improving the possibility of detecting the ROI. The synthesized RGB image is analyzed using a fine-tuned RetinaNet model to detect the ROI. The performance of the RetinaNet model was assessed and compared with four leading deep learning object detection models, using a dataset of 380 breast ultrasound images. The results demonstrate that the RetinaNet model outperformed the other models, with a failure rate of 7%, a mean recall value of 89%, a mean precision value of 89%, and a mean F1-score value of 88%. These findings suggest that the RetinaNet model provides a feasible and effective approach for detecting the ROIs in breast ultrasound images.