A review of performance of recent YOLO models on cholecystectomy tool detection

Recent years have seen a drastic advancement in the performance of deep learning algorithms for object detection. On the one hand, their prediction accuracy has improved tremendously. On the other hand, the speed with which they detect objects has also improved markedly. This is also coupled with the substantial improvement in the computational hardware. Graphics processing hardware is not only commonplace, it also keeps on taking strides in terms of the number of computational cores per chip, as well as the performance improvement per core. All of these advancements have paved the way for challenging opportunities in various application domains to benefit from.

In this paper, we have compared the performance of various variants of the famous ”you only look once” object detection algorithm for detecting computer-aided laparoscopy tools used in cholecystectomy procedures. We trained various state-of-the-art models using a well-known benchmark dataset. Our results for some models are particularly astounding. We report both on the accuracy and computational efficiency of different models. In terms of accuracy, we have chosen mean average precision as the main metric. Mean average precision is reported for each model against each tool. For the best model, we report several metrics that include precision, recall, and mean average precision. Apart from this, we have also plotted the confusion matrices for all the models on the unseen test data. For computational efficiency, we report the time it takes by each model during various inferential phases.