Background: For electrosurgery of the prostate, which relies on surveillance screens for real-time operations, manual remains the primary method for prostate capsula identification, rapid and accurate detection becomes urgency.We aimed to develop a deep learning method for detecting prostate capsula using endoscopic optical images.
Methods: Firstly, the SimAM residual attention fusion module is used to enhance the feature extraction ability of texture and detail informations. Secondly, the enhanced details information at the lower level is transferred to the higher level in a hierarchical manner to facilitate the extraction of semantic information. Finally, based on SimAM residual attention and forward hierarchical feature-by-feature fusion, an improved single-shot multibox detector model based on 3D unparametric attention fusion is proposed.
Results: A forward feature-by-feature hierarchical fusion network based on the 3D residual attention mechanism is proposed for the fast detection of the prostate capsula. The proposed network can better extract the regional features of the image, the extracted network features retain the spatial structure of the image, and realize the rapid detection of medical images.
Conclusions: The speed of the proposed model can reach 0.014ms on NVIDIA RTX 2060, which realizes the rapid detection. The model AFFSSD composed of unparametric attention fusion and progressive fusion of forward features can achieve 83.12% detection precision. Compared with Faster R-CNN (ZF, VGG16, ResNet 50), SSD(VGG16, ResNet 101), EfficientDet (D0-D7), FoveaBox, TOOD, YOLOv4, Sparse R-CNN, OWOD, .R-FCN(ResNet-50), FSSD(VGG16), the proposed method AFFSSD had the highest mAP and faster speed, only lower than YOLOv7.