Abstract
Background
Accurate identification and localization of prostate zones in magnetic resonance (MR) images are essential for clinical diagnosis and treatment planning. However, convolutional object detection models like YOLO often struggle to capture the complex geometric features of the prostate.
Objective
To enhance the detection and segmentation performance of prostate MR images by addressing limitations in spatial feature extraction and static focusing mechanisms present in conventional YOLO models.
Methods
We propose YOLO-D, an enhanced YOLOv8-based model integrating a Deformable Convolution (DConv) module to better capture fine-grained image details and improve geometric adaptability. Additionally, the Wise-IoU loss function is employed to introduce a dynamic and non-monotonic focusing mechanism, effectively reducing inter-class interference and enhancing localization accuracy.
Results
YOLO-D was evaluated on the publicly available ProstateX dataset using precision, recall, average precision (AP), and F1 score as evaluation metrics. For detection, it achieved 93.4% precision, 91.2% recall, 94.7% AP, and an F1 score of 0.922. For segmentation, YOLO-D achieved 90.7% precision, 88.6% recall, 91.1% AP, and an F1 score of 0.897—consistently outperforming the baseline YOLOv8.
Conclusions
By incorporating DConv and Wise-IoU, YOLO-D offers a robust and efficient solution for automatic prostate zone analysis, with promising potential in real-time clinical imaging applications.
Get full access to this article
View all access options for this article.