Visual Modalities Based Multimodal Fusion for Surgical Phase Recognition

Abstract

We propose visual modalities-based multimodal fusion for surgical phase recognition to overcome the limitation of the diversity of information such as the presence of tools. Through the proposed methods, we extracted a visual kinematics-based index related to the usage of tools such as movement and the relation between tools in surgery. In addition, we improved recognition performance using the effective fusion method which is fusing CNN-based visual feature and visual kinematics-based index. The visual kinematics-based index is helpful for understanding the surgical procedure as the information related to the interaction between tools. Furthermore, these indices can be extracted in any environment unlike kinematics in robotic surgery. The proposed methodology was applied to two multimodal datasets to verify that it can help to improve recognition performance in clinical environments.