Research On Endoscopic Image Smoke Removal Method Based On Attentional Multilevel Feature Learning

As modern medical tools continue to develop,electrosurgical devices such as ultrasonic knives and electrocoagulation forceps are increasingly utilized in endoscopic surgery.While these devices have improved the convenience and effectiveness of endoscopic surgery,the surgical smokes generated during the process can pose a potential hazard to surgeons.The resulting smoke can obscure visibility of the surgical field,making its removal essential for safe and efficient performance of the procedure.Although mechanical smoke evacuation devices are currently used for this purpose,they may negatively impact surgical efficiency,harm patients,and lead to increased treatment costs.In recent years,image processing technologies and artificial intelligence have advanced significantly,resulting in the development of medical devices based on deep learning models that have been implemented in health care.This thesis proposes two methods for removing smoke from endoscopic surgical images using generative adversarial networks.Specifically,this study divides the surgical smoke removal task into two subtasks: smoke detection,which accurately identifies smoke and assists in the smoke removal task,and smoke removal,which uses a generative adversarial learning algorithm to achieve surgical smoke removal.To train the model,the study employs a powerful smoke rendering engine to simulate the generation of endoscopic surgical smoke,which is then randomly rendered onto the endoscopic surgical images to form paired synthetic datasets.By training the generative adversarial network on these datasets,the method effectively removes surgical smoke from endoscopic surgical images.The thesis works as follows.(1)In this study,a generative adversarial network model(MSF-GAN)based on multi-level smoke feature learning is constructed for surgical smoke removal.This technique adaptively learns non-homogeneous smoke features using specific branches and integrates global features to preserve the semantic and textural information of pyramidal connections.To improve the smoke removal capability of the model,we synthesized a paired dataset of smokeless and smoky data using Blender,a 3D smoke rendering engine,for the training of the smoke removal model.We conducted various experiments on an endoscopic surgical image dataset,and the results showed that our proposed model effectively removed endoscopic surgical smoke.Furthermore,ablation experiments demonstrated that the multilevel smoke feature learning strategy we proposed significantly improved the smoke removal performance of the smoke removal model for endoscopic images.(2)In this study,an attention-aware generative adversarial network model(MARS-GAN)based on multi-level feature learning is constructed for surgical smoke detection and removal.The MSF-GAN model was used as a basis but resulted in color distortion or artifacts in smoke-free areas.Therefore,a smoke detection network was added,which combined a smoke segmentation network with a dark channel a priori module to provide pixel-level measurements for focusing on smoke features while preserving smoke-free details.The multi-task learning strategy included adversarial loss,cyclic loss,smoke perception loss,dark channel prior loss,and contrast enhancement loss for model optimization.The proposed model was tested on synthetic and real endoscopic surgical image datasets and validated on the publicly available Cholec80 dataset,which outperformed 14 existing smoke or haze removal methods.Ablation experiments on different modules and loss functions showed that the proposed approach facilitated surgical smoke removal by the model.In this thesis,we propose an attention-aware generative adversarial network for surgical smoke removal in endoscopic images,based on multilevel feature learning.The method effectively reduces surgical smoke and restores true color in the abdominal cavity,with both research and clinical significance for endoscopic surgery.