A Deep Dive Into Capsulorhexis Segmentation: From Dataset Creation to SAM Fine-tuning

Capsulorhexis is the most fateful phase of cataract surgery. The ability to automatically segment its main regions, and thereby, extract information can lead to image-guided surgery. Image-guided surgery has a wide range of applications that contribute to improved surgical outcomes and lower clinical risks. Despite the significant importance of capsulorhexis segmentation, a dedicated dataset focusing on this phase of cataract surgery is currently unavailable. This paper bridges this gap by creating a comprehensive dataset, which is named ARAS-CaSe, developed exclusively for capsulorhexis segmentation. ARAS-CaSe wide variety dataset is an invaluable resource for training computer vision models and achieving advancements in surgical skill assessment. Furthermore, certain state-of-the-art segmentation models were trained and fine-tuned on this dataset. Among the evaluated models, the SAM model had the highest level of performance, with an Intersection over Union (IoU) score of 91% and a Dice coefficient of 95.11%. To further analyze the models, we have performed a comparative analysis to find the efficient loss function for the U-Net segmentation model in the eye surgery domain. This study was conducted independently for each dataset class, and the presented results show that the Lovász-Softmax loss function will produce the best outcomes for capsulorhexis segmentation. ARAS-CaSe dataset, saved models, and codes of this research will be available by submitting a request through this website (aras.kntu. ac.ir/ai).