A Novel Data Augmentation Method for Robotic Surgical Instrument Small Part Segmentation in Complex Scenes.

Accurate and robust segmentation of surgical instrument parts is urgently required by AI enhanced intelligent surgery and automatic surgery skill evaluation. However, it is still a challenging problem due to the multiple small areas of frontal clasper and wrist parts, complicated object occlusion and light variation issues. Increasing high quality labeled data is effective to improve the performance but labor-costing. A novel instrument data augmentation method is proposed to generate high quality training data through planar instrument replacement and orientation adjustment. Moreover, image repainting methods are utilized to recover the occluded parts by instruments and generate clean tissue images. Then joint angle reconfigured instruments are placed on recovered clean tissue images to generate simulated surgery images, and we control the occlusion rates and overlapping areas by which high quality instrument manipulation augmented training data are generated automatically. Detailed experiment results show our proposed data augmentation method make existing methods achieve better segmentation accuracy and robustness of wrist and clasper part, especially the fault detection errors decreased significantly in instrument overlapping and complicated lighting conditions.