Optimized Convolutional Neural Network for Smoke Detection in Laparoscopic Surgery Using Multi-Modal Feature Fusion

In minimally invasive surgeries, particularly laparoscopic procedures, smoke generated from electrocauterization can severely obstruct the surgeon's view, leading to potential complications and extended operation times. This visual impairment poses a significant challenge, as it can delay critical decisions and increase the risk of errors. Addressing this issue, we propose a sophisticated convolutional neural network (CNN) model tailored for detecting smoke in laparoscopic images. The proposed model leverages a combination of spatial image data and extracted numerical features, such as Normalized-RGB, Wavelet energy, GLCM texture, HSV color space, Optical Flow, and Fog Area, to achieve high accuracy in identifying smoke presence. The dataset used in this study comprises frames and video clips from ten robot-assisted laparoscopic hysterectomy procedures, meticulously annotated to highlight regions containing smoke. The CNN architecture integrates both image-based and feature-based inputs, passing them through multiple convolutional layers to capture intricate patterns and textures. The final model, optimized with the Adam optimizer and a learning rate of 0.0001, demonstrates a significant improvement in smoke detection, achieving an accuracy of 92.6%. This advancement holds promise for enhancing real-time surgical assistance systems, potentially improving surgical outcomes by reducing the time and risks associated with smoke-obstructed views.