An Experimental Study of nmODE in Recognizing Endoscopic Submucosal Dissection Workflow

Endoscopic Submucosal Dissection (ESD) is an innovative and minimally invasive endoscopic surgical technique that is currently considered the preferred therapeutic approach for addressing early-stage gastrointestinal cancers. Accurate identification of the ESD surgical workflow is crucial for education and clinical practice. Due to the complexity of ESD surgical procedures, directly applying existing surgical workflow recognition methods to ESD data might lead to compromised performance. Thus, a more robust approach is required to extract surgical features from ESD videos. The recently introduced neural memory Ordinary Differential Equations (nmODEs) have demonstrated substantial potential in learning nonlinear rep-resentations. To fully harness its remarkable representational capabilities, we introduces a nmODE-based two-stage algorithm for recognizing ESD workflows. In the first stage, a nmODE-ResNet model is proposed to extract discriminative visual features from ESD videos. In the second stage, four sequential models, including LSTM, GRU, TeCNO, and Transformer, are employed to aggregate the temporal information of the ESD procedure. The proposed algorithm is extensively tested on a dataset comprising 62 ESD surgical videos. The results indicate that the proposed nmODE-ResNet model achieves superior performance when compared with other competitive models.