Real-Time Intelligence using Deep Learning for Skill Enhancement in Capsulorhexis Cataract Surgery

This research focuses on creating a guidance map for computer-assisted cataract surgeries. The main focus is to improve surgery duration and assist young practitioners while performing cataract surgeries with the capsulorhexis method. The research features two phases, namely, the Incision Phase and the Capsulorhexis Phase among thirteen different phases of cataract surgery. This research provides methods for real-time phase recognition and segmentation. The study uses ResNet-18 for phase recognition to identify the Incision Phase and the Capsulorhexis Phase. The model was able to achieve an accuracy of 94.42% and F1 Score of 0.9452. UNet with Guidance Filter Module (GFM) is used to segment the cornea and the main incision knife. The UNet model produced masks with a mean IoU of 0.9315, mean Dice of 0.9642 with an inference time of 6.2ms for incision knife segmentation. For the task of cornea segmentation, the model produced results with an inference time of 6.5ms with mean IoU score of 0.9450 and mean Dice score of 0.9716. Using the extracted features and the recognized surgical phase, a guidance map is plotted for each frame of the video. These frames are rendered at 30 Frames Per Second (FPS) to create a video. The Incision Phase map provides a guide for the main incision and the side ports along with horizontal depth tracking of the main incision. The capsulorhexis guidance map provides a guide along the rhexis circle (circle along which capsule is torn).