2.4. Instance Segmentation Framework

The instance segmentation framework, Mask R-CNN [39], consisted of sequence algorithms of region proposal networks (RPN) [41] for detecting SIs and semantic segmentation networks (Figure 2a). Unlike semantic segmentation methods applied to robotic surgery images [42,43], the proposed instance segmentation method separates occluded instruments during the first stage of RPN, followed by application of a semantic segmentation network during the next stage. Surgical instruments that were only partially visible on the screen were defined as indistinguishable. Therefore, the datasets were trained using a binary cross-entropy loss to approach a binary rather than a multi-class classification task (Supplementary Figure S1).