Residual Network (ResNet-18) for Laparoscopic Image Distortion Classification

EasyChair Preprint 11575

Abstract

Diminished laparoscopic video quality directly affects a surgeon's visibility and can compromise the outcomes of computational tasks in robot-assisted surgery. To address this challenge, numerous solutions have been proposed based on the detection and classification of laparoscopic video distortions. In this work, we propose a method based on Residual networks (ResNet18) for the automatic detection and classification of noise ‘NO’, smoke ‘SM’, uneven illumination ‘UI’, defocus blur ‘DB’, and motion blur ‘MB’ in laparoscopic videos. We have obtained an accuracy of 98.75% for training and 97.97% for validation. The high accuracy scores across the classes emphasize the model's capability to generalize well and make accurate predictions.

Keyphrases: Laparoscopic video, deep learning, distortion classification

@booklet{EasyChair:11575,
  author    = {Mohamed Belmokeddem and Kamila Khemis and Salim Loudjedi},
  title     = {Residual Network (ResNet-18) for Laparoscopic Image Distortion Classification},
  howpublished = {EasyChair Preprint 11575},
  year      = {EasyChair, 2023}}