Deep Learning Approach for Bubble Segmentation from Hysteroscopic Images

Gas embolism is a potentially serious complication of hysteroscopic surgery. It is particularly necessary to monitor bubble parameters in hysteroscopic images by computer vision method for helping develop automatic bubble removal devices. In this work, a framework combining a deep edge-aware network and marker-controlled watershed algorithm is presented to extract bubble parameters from hysteroscopy images. The proposed edge-aware network consists of an encoder-decoder architecture for bubble segmentation and a contour branch which is supervised by edge losses. The post-processing method based on marker-controlled watershed algorithm is used to further separate bubble instances and calculate size distribution. Extensive experiments substantiate that the proposed model achieves better performance than some typical segmentation methods. Accuracy, sensitivity, precision, Dice score, and mean intersection over union (mean IoU) obtained for the proposed edge-aware network are observed as 0.859 ± 0.017, 0.868 ± 0.019, 0.955 ± 0.005, 0.862 ± 0.005, and 0.758 ± 0.007, respectively. This work provides a valuable reference for automatic bubble removal devices in hysteroscopic surgery.