Evaluating pre-processing and deep learning methods in medical imaging: Combined effectiveness across multiple modalities

Abstract

Medical imaging is critical in modern healthcare for accurately detecting and diagnosing various medical conditions. Advanced computational techniques, particularly preprocessing methods and deep learning models, have demonstrated significant potential for improving medical image analysis. However, determining the optimal combination of these techniques across different types of medical images remains a challenge. Using empirical experiments, this evaluation research investigates the effectiveness of five popular pairs of preprocessing techniques combined with five widely used deep learning models. Preprocessing methods evaluated include CLAHE + Butterworth, DWT + Threshold, CLAHE + median filter, Median-Mean Hybrid Filter, and Unsharp Masking + Bilateral Filter, concatenated with deep learning models: EfficiencyNet-B4, ResNet-50, DenseNet-169, VGG16 and MobileNetV2. The performance of these combinations was evaluated through experiments carried out on eight diverse and commonly used datasets encompassing various medical imaging modalities. These datasets include two X-ray collections: the COVID-19 Pneumonia Normal Chest PA Dataset and the Osteoporosis Knee X-ray Dataset; two CT scan datasets: the Chest CT-Scan Images Dataset and the Brain Stroke CT Image Dataset; two MRI datasets: the Breast Cancer Patients MRI and the Brain Tumor MRI Dataset; and two ultrasound datasets: the Ultrasound Breast Images for Breast Cancer and the MT Small Dataset. Our findings show that the Median-Mean Hybrid Filter and Unsharp Masking + Bilateral Filter are the most effective preprocessing methods, achieving an efficiency rate of 87.5%. Among the deep learning models, EfficiencyNet-B4 and MobileNetV2 are the highest performing models with an efficiency ratio of 75%, with MobileNetV2 providing up to 34% shorter runtime compared to other models. This study provides a thorough evaluation of the performance of different preprocessing methods and deep learning algorithms across commonly used medical imaging modalities. Presenting empirical results from our experiments offers practical insights into choosing the most suitable preprocessing techniques and deep learning models for various types of medical images. These findings are intended to support improvements in diagnostic accuracy and efficiency in medical imaging, offering a valuable reference for enhancing image-based diagnostic processes.