Prostate Localization in 2D Sequence MR with Fusion of Center Position Prior and Sequence Correlation

A prostate organ localization algorithm based on the state-of-art object detection framework Faster R-CNN is proposed in this paper for Magnetic Resonance (MR) slice sequence. Using the prior information of the central position of the organ in the image, ResNet-50 with spatial attention mechanism is introduced as the network's feature extraction module to enhance the sensitivity of the network to the spatial location features. In addition, inspired by the correlation between neighboring slice images in the position and morphological size of the organ, spatial curve fitting of key points of the object bounding boxes based on the sequence direction is applied to further improve the detection performance of the algorithm. Compared with the original Faster R-CNN framework, the algorithm we proposed has achieved better performance of prostate localization on the PROMISE12 dataset, which is mainly reflected in the area recall rate and localization success rate of the organs increased by 7.1% and 6.5%, respectively. It establishes a good foundation for subsequent medical image processing tasks such as organ segmentation and lesion detection.