BLVQE: Blind Laparoscopic Video Quality Evaluator using spatio-temporal interdependency and textural cues

Abstract

Quality assessment of laparoscopic videos is critical for ensuring accurate diagnostics and surgical precision. Traditional quality assessment methods typically focus on either spatial or textural features independently, limiting their effectiveness in handling composite distortions like motion blur, noise, defocus blur, uneven illumination, and smoke. To address this, leveraging spatio-temporal interdependencies and textural features offers a more comprehensive approach in replicating the human visual system to improve the robustness of video quality assessment. This work introduces Blind Laparoscopic Video Quality Evaluator (BLVQE) that models the statistical interdependencies between spatial, temporal and texture features. For this, laparoscopic videos obtained from a public database are used to estimate the Luminance and motion vector maps, which are then analyzed using bivariate generalized Gaussian distribution to capture spatio-temporal interdependency. Scene texture complexity is further quantified using statistical energy measures. These feature vectors are used for end-to-end training of an LSTM framework for frame quality predictions. The training and validation loss curves of the model saturate around 50 epochs, indicating prediction proficiency. BLVQE predictions show a high correlation with subjective scores exhibiting robust and competitive performance against other state-of-the-art methods. Ablation studies highlight the contribution of individual feature elements, confirming the superiority of the selected features. These findings enhance the understanding of the spatial, temporal and textural variations that influence video quality and highlight the potential of joint dependencies in accurately estimating the diagnostic quality of laparoscopic videos.