Sutures and Landmarks Joint Detection Method Based on Convolutional Neural Network for Rat Stereotactic Surgery

With the advantages of manual error reduction in manipulating instruments, surgical robots have shown great value in stereotactic surgery. Most of them require doctors to locate insertion sites manually, which introduces empirical errors. Some researchers have discussed deep-learning based methods to automatically segment medical images and locate targets. However, the existing studies mainly focus on the general-sized objects, while the landmarks of rat skull are a few pixels in size. Besides, the serious image noise produces further challenges. To solve the problems, we propose a sutures and landmarks joint detection method, which contains a two-branch framework, a hybrid loss and a class-specific pretraining method. The two-branch framework contains a shared encoder and two parallel decoders connected by feature fusion modules. The hybrid loss includes an intersection of union loss between different classes of landmarks and a contrastive feature loss to make the landmarks decoder extract more discriminative feature. The class-specific pretraining method supervise the two parallel decoders with boundary label and object label respectively, which makes the decoders learn related decoding processes in pretraining. A novel dataset containing sixty skull images of twenty Sprague Dawley rats is constructed for training and test. Experimental results show that the proposed method achieves 0.01 mean absolute error in sutures segmentation and over 80% correct rate in landmarks location.