Soft Annotations versus Pixel-Based Segmentation Masks of Prostate Anatomies: The Effect of Annotation Type on...

Prostate cancer detection and segmentation is a vastly challenging task. Factors such as the low MRI contrast, small lesion size, high inter-observer bias of the ground truth, different acquisition protocols, and scanners with different specifications pose significant difficulties in deep learning segmentation of the region of interest. Additionally, magnetic resonance imaging is used as the baseline decision support image modality, which contributes to the parameter variability of the examined input data. Detecting the region of interest with bounding boxes can be used as a less precise but highly efficient alternative for deep learning-based analyses. In this study, several deep learning architecture variations of the YOLOv8 were used to detect prostate glands and potential neoplasms. The gland detection model achieved a mAP of 95.8±1% on the unseen testing sets, while the best prostate lesion detection model yielded a mAP of 41.5±7%. Additionally, an analysis of the impact of bounding boxes compared to pixel-based annotations on imaging features was carried out. The effect of image quality on gland detection with deep learning-based models was assessed.Clinical Relevance- Annotating large datasets, especially in oncology, is a big barrier for many understaffed research and medical groups. Additionally, the high inter-observer variability can be a significant factor that affects the quality of machine learning models. The proposed cascaded deep learning analysis for detecting prostate lesions, which is a common malignancy in men, can alleviate some of these concerns and can accelerate the tedious and time-consuming lesion annotation process for large multi-modal MRI datasets.