Quality Evaluation of Laparoscopic Videos Using the Interdependency of Luminance and Texture Maps

The perceived quality of the video during minimally invasive surgeries greatly impacts the success of these medical procedures. Consequently, there is high demand for varied methods aimed at quality assessment (QA) of laparoscopic videos. In this work, a video quality prediction model is proposed to explore the statistical dependency between the subband coefficients of luminance and texture components of motion blurred laparoscopic videos. For this, the publicly available motion blurred videos from the Laparoscopic Video Quality (LVQ) database is used. The luminance map and gray-level co-occurrence matrix (GLCM) texture map are computed for each video frame and decomposed using steerable pyramid decomposition method. The decomposed subband coefficients are modelled using Bivariate Generalised Gaussian Distribution (BGGD) and the distribution depicts sharp peaks and long tails. The BGGD $(\alpha,\beta)$ parameters are estimated for each subband level and they are average pooled to estimate the frame-level values. The videos of the LVQ database is randomly divided into 66:34 training-to-testing ratio to perform supervised non-linear regression using Adaptive Neuro Fuzzy Inference System (ANFIS) network. In training phase, frame-level BGGD parameters are used as inputs along with expert and non-expert Mean Opinion Score (MOS) values as training labels for each video. In testing phase, the quality score of each test video is predicted. The performance of the proposed model is compared to other 2D image QA (IQA) and video QA (VQA) algorithms. The Pearson’s Linear Correlation Coefficient (PLCC) and Root Mean Square Error (RMSE) performance metrics are used to evaluate the prediction efficiency of the model. It is observed that the proposed model performs prediction with a PLCC value of 0.9814 for expert training labels. When non-expert training labels are used a PLCC value of 0.9635 is obtained. The model also demonstrates an RMSE value of 0.3939 when expert MOS training labels are used, and an RMSE value of 0.4639 when non-expert MOS training labels are used. The higher correlation scores and lower RMSE values indicate that the model can be used as a VQA technique for analysis of laparoscopic videos affected by blur distortions due to motion.