Systematic investigation of pre-processing and feature extraction techniques in medical image analysis

Abstract As medical imagery remains the cornerstone of diagnosis, the success of complex classification tasks depends on high-quality data and diagnostic features. This study evaluates image pre-processing and feature extraction to optimize model performance. Our investigation assesses how variations in pre-processing and feature extraction impact classification efficiency across three imaging modalities: radiology (chest X-ray images), pathology (H&E (Hematoxylin and Eosin)-stained patches), and ophthalmology (OCT (Optical Coherence Tomography) scans). The experimental framework incorporates nine pre-processing techniques, adjustment (brightness, contrast, and histogram equalization), filtering (mean, median, and Gaussian), and three normalization schemes, systematically combined for each modality. Features were extracted using five pre-trained architectures (VGG16, ResNet50, DenseNet121, MobileNetV2, and InceptionV3) and classified via a PCA-LDA (Principal Component Analysis-Linear Discriminant Analysis) pipeline. Performance was evaluated using the mean sensitivity score. Mean sensitivity scores improved significantly: H&E-stained images increased from 74·9% to 96·95%, chest X-rays from 89·9% to 96·65%, and OCT scans from 82·4% to 98·8%. No single pre-processing configuration consistently dominated; optimal settings varied by modality, confirming the absence of a universal solution. VGG16 and DenseNet121 exhibited the greatest robustness. Implementing at least one pre-processing step combined with a robust DL feature extractor can substantially improve model efficacy in diagnostic detection tasks.