Dynamic data balancing strategy-based Xception-dual-channel LSTM model for laparoscopic cholecystectomy phase...

Area of Science:

• * Medical image analysis
• * Surgical phase recognition
• * Deep learning for surgical video analysis

Background:

• * Laparoscopic cholecystectomy phase recognition is crucial for surgical safety and training.
• * Existing models struggle with temporal feature learning and class imbalance in surgical datasets.
• * Enhancing temporal feature extraction and data balancing is key to improving model accuracy.

Purpose of the Study:

• * To develop an advanced deep learning model for laparoscopic cholecystectomy phase recognition.
• * To improve temporal feature learning capabilities and address data class imbalance.
• * To propose a novel Xception-dual-channel LSTM fusion model with a dynamic data balancing strategy.

Main Methods:

• * Employed an Xception model with depthwise separable convolutions for frame-by-frame visual feature extraction.
• * Utilized a dual-channel LSTM network (temporal mapping bidirectional LSTM and sequence embedding LSTM) to model temporal dependencies.
• * Implemented a dynamic data balancing strategy to adjust undersampling rates and mitigate biased learning.

Main Results:

• * The proposed Xception-dual-channel LSTM model outperformed traditional single-channel LSTM models on the Cholec80 dataset.
• * F1-scores for all surgical phases were improved compared to models without dynamic data balancing.
• * Experimental results confirmed the model's effectiveness in handling temporal features and class imbalance.

Conclusions:

• * The dynamic data balancing strategy effectively alleviates class imbalance issues in surgical phase recognition.
• * The Xception-dual-channel LSTM fusion model demonstrates superior performance in extracting temporal features.
• * The proposed approach significantly enhances the overall detection performance for laparoscopic cholecystectomy phases.