Automated segmentation by deep learning of loose connective tissue fibers to define safe dissection planes in robot-assisted gastrectomy

Abstract

The prediction of anatomical structures within the surgical field by artificial intelligence (AI) is expected to support surgeons' experience and cognitive skills. We aimed to develop a deep-learning model to automatically segment loose connective tissue fibers (LCTFs) that define a safe dissection plane. The annotation was performed on video frames capturing a robot-assisted gastrectomy performed by trained surgeons. A deep-learning model based on U-net was developed to output segmentation results. Twenty randomly sampled frames were provided to evaluate model performance by comparing Recall and F1/Dice scores with a ground truth and with a two-item questionnaire on sensitivity and misrecognition that was completed by 20 surgeons. The model produced high Recall scores (mean 0.606, maximum 0.861). Mean F1/Dice scores reached 0.549 (range 0.335-0.691), showing acceptable spatial overlap of the objects. Surgeon evaluators gave a mean sensitivity score of 3.52 (with 88.0% assigning the highest score of 4; range 2.45-3.95). The mean misrecognition score was a low 0.14 (range 0-0.7), indicating very few acknowledged over-detection failures. Thus, AI can be trained to predict fine, difficult-to-discern anatomical structures at a level convincing to expert surgeons. This technology may help reduce adverse events by determining safe dissection planes.

Figure 1

Deep learning algorithm and the AI model developed in this study. (a) The deep-learning architecture implementing U-Net. Conv, convolution; concat, concatenation. (b) Development and performance evaluation of the AI model. MR misrecognition.

Figure 2

The questionnaire for qualitative evaluation of the AI’s segmentation performance completed by expert surgeons.

Figure 3

Comparison of segmentation performance at different stages in deep learning. (a) An original frame. CHA common hepatic artery, F fat tissue, LN lymph node; *, nerve. (b) Magnified view of the square in A showing prediction of loose connective-tissue fibers (LCTFs) highlighted in turquoise by the prototype AI model. White circle indicates an area of over-detection. (c) Prediction by the latest AI model. Arrows indicate LCTFs that could not be detected by the prototype AI model.

Figure 4

Relations between computed performance metrics and qualitative scores. (a) A mosaic diagram showing the distribution of all scores assigned by 20 evaluators to 20 randomly sampled frames. Blue, light blue, and gray panels respectively represent scores of 4, 3, and 2 for Question 1 (see Fig. 2). Vertical and horizontal axes respectively represent the proportion of scores assigned to Questions 1 and 2. Values in the rectangles represent the ratio of each category against the total. There were no scores below 1 for Question 1 and no scores above 3 for Question 2. (b) Scatter plot showing the relation between sensitivity and misrecognition (MR) scores for each frame. Blue area is the confidence ellipse, representing the area of 95% probability that the plots exist. (c) Scatter plot showing the relation between sensitivity and Recall scores. The correlation coefficient was 0.733 and the 95% confidence interval was 0.430–0.887. Blue line represents the regression formula, calculated as Y = 2.302 + 2.001X. Y sensitivity score, X Recall score.

Figure 5

AI prediction results for (a) frame 6 with the highest sensitivity score and (b) frame 19 with the lowest sensitivity score. The area surrounded by the broken line is an under-detection area.

Figure 6

Examples where the AI misrecognized (a) gauze mesh fiber, (b) fine grooves at the tips of forceps, and (c) minor halation of fat or blood surfaces as loose connective tissue.