Unsupervised smoke to desmoked laparoscopic surgery images using contrast driven Cyclic-DesmokeGAN

In laparoscopic surgery, energized dissecting devices and laser ablation causes smoke, which degrades the visual quality of the operative field. This paper proposes an unsupervised approach to desmoke laparoscopic images called Cyclic-DesmokeGAN. In the generator, multi-scale residual blocks help to alleviate the smoke component at multiple scales, while refinement module helps to obtain desmoked images with sharper boundaries. As the presence of smoke degrades contrast and fine structure, the proposed method utilizes high boost filtered image at each encoder layer. The contrast loss improves overall contrast, thereby reducing the smoke, while Unsharp Regularization loss helps to stabilize the network. The proposed Cyclic-DesmokeGAN is tested on 200 smoke images obtained from Cholec80 dataset consisting of videos of cholecystectomy surgeries. The results depict effectiveness, as proposed approach achieved 3.47±0.09 Contrast-Distorted Images Quality, 4.15±0.74 Naturalness Image Quality Evaluator, and 0.23±0.00 Fog Aware Density Evaluator, these indexes are best in comparison to other state-of-the-art methods.