Combinational Distortion Assessment of Laparoscopic Videos using Deep EfficientNet Model

Laparoscopic images are susceptible to several distortions making it challenging for surgeons to have a clear visibility of images/videos during surgery. Distortions in images might lead to misdiagnosis, leads to adverse consequences. An objective technique for assessing the laparoscopic image quality is required in order to identify the type of distortion. Most of the laparoscopic image datasets with distortion labels are not publicly available and are limited to certain distortions. Laparoscopic Image Quality Assessment (LAPIQA) dataset is created with the necessary labels specifically for working with this model. It comprises of three distinct distortions blur, smoke/fog and uneven illumination each of which has five levels and can be either a single distortion or combination of distortions. A deep learning based multi-label distortion classification model is implemented to identify the various types of distortions. Modified EfficientNetBO is used with added custom layers and sigmoid activation for the model. The model is trained with both hinge loss and binary cross entropy (BCE) and validated using metrics such as accuracy, loss, precision, recall and f1-score. With respect to accuracy, model yields an average training accuracy of 96.51%, with a training loss of O.55, validation accuracy of 95.85%, with a validation loss of O.57 and testing accuracy of 96.12% with a loss of O.71. The proposed classification approach effectively identifies and evaluates both single and combined distortions in laparoscopic images.