Use of Deep Learning for Liver Segmentation during Laparoscopic Cholecystectomy

During laparoscopic cholecystectomy, Indocyanine Green (ICG) injection and near-infrared fluorescence (NIRF) imaging techniques are employed to detect the boundaries of the liver, vessels, gallbladder, and biliary structure by changing their color to green. It takes time for ICG to flow inside vessels before liver visualization. Moreover, this technique needs ICG injection for each visualization. The current study proposes a deep learning-based method that segments the liver in realtime. The public dataset CholecSeg8k is employed for network training, validation, and testing. A private dataset from KPJ Damansara, Malaysia, is also used for testing. The public Python library, segmentation models, is utilized for liver segmentation implementation. The U-Net architecture combined with the SE-ResNet152 backbone produced the most accurate liver segmentation result among the experiments. In the first top result, the evaluation of the model on the test set achieved a mean intersection over union (IoU) score of 0.96064 and a mean F-score of 0.97953. During laparoscopic cholecystectomy, the automated liver segmentation method may be considered an alternative to conventional techniques relying on ICG-NIRF. In the future, exploring a robust network model to improve the result for the private dataset will be investigated.