A Systematic Investigation of Image Pre-Processing on Image Classification

Abstract

AI-powered image analysis is a transformative technology with immense potential to enhance diagnostics and patient care. Accurate medical image assessment plays a crucial role in disease detection and treatment planning, yet challenges arise due to noise and visual variations in medical imaging. Image pre-processing is a key solution to address these challenges, and while widely used, there is a lack of studies on its effectiveness. Recognizing this gap, our research aims to contribute insights to this scientific scope. This research specifically delves into the impact of pre-processing on the binary classification model performance, rather than model and hyperparameter optimization. We deliberately selected a limited yet comprehensive subset of methods and datasets; H&E-stained tissue, chest X-ray, and retina OCT images were chosen to ensure the generalizability of our findings. Analysis revealed that implementing a pre-processing significantly improved mean sensitivity in the binary classification models: from 0.87 to 0.97 for H&E-stained tissue, 0.92 to 0.96 for chest X-rays, and 0.96 to 0.99 for Retina OCT images. Two different sequences for applying pre-processing steps were explored, with minimal effect observed in the altered sequences, indicating consistent improvement regardless of the chosen sequence. We investigated the pre-processing steps employed in the 40 of the best-performing and worst-performing models, determined by the higher and lower mean sensitivities. We have uncovered that the pre-processing steps of the best-performing models displayed only minimal similarities, except for the pooling mode. This observation also applied to the worst-performing models with lower sensitivity.