Deep Learning-based Intraoperative Video Analysis for Cataract Surgery Instrument Identification

Surgical instrument detection and classification is a critical task for enhancing surgical procedures monitoring, assisting surgical operations, supporting medical education, and enabling the development of intelligent surgical systems. However, there are a few challenges in this domain. The foremost concern is the impact of varying background conditions. Additionally, class imbalance presents another challenge, potentially leading to biased classification results. To solve these challenges, this study proposes a deep learning-based system consisting of two key components: an attention region detection module and a ResNet50 classification model. The attention region detection employs an optical flow-based method to incorporate both temporal and spatial information from the surgical video so that critical attention regions covering surgical instruments are identified. Our experimental results show that the classification accuracy can be improved from 58.7% to 81.9% by using the attention region detection component. To deal with the challenge of class imbalance, we use focal loss and interleaved sampling strategy as solutions. Interleaved sampling uses both the spatial and temporal information of surgical videos to balance the number of samples across different instrument classes, through which some scarce surgical instrument classes are expanded, thus preventing biased learning of the model. And the validation accuracy on the balanced dataset achieves 87.1%. This study demonstrates the effectiveness of deep learning techniques in addressing challenges in cataract surgery video analysis.