Enhancing Medical Image Classification through Transfer Learning and CLAHE Optimization

What is it about?

This study investigates how to improve brain medical image classification by combining transfer learning with Contrast Limited Adaptive Histogram Equalization (CLAHE). Medical image datasets often suffer from limited size and variable image quality, which can negatively affect deep learning performance and lead to overfitting. To address these challenges, we applied CLAHE as a preprocessing technique to enhance local contrast and improve feature visibility in brain MRI images. We then evaluated the impact of this enhancement on classification performance using pre-trained deep learning models (VGG16, VGG19, and ResNet50) under different conditions, including with and without data augmentation. The study systematically compares classification accuracy, convergence behavior, and overfitting tendencies across four experimental settings: original images, CLAHE-enhanced images, and each condition with and without data augmentation. Our findings demonstrate that CLAHE significantly improves model stability and classification accuracy, particularly when combined with data augmentation. The results highlight the importance of optimized preprocessing strategies in medical imaging workflows and provide practical guidance for improving diagnostic AI systems when working with limited datasets.

Featured Image

Why is it important?

Accurate medical image classification is essential for supporting early diagnosis, treatment planning, and clinical decision-making, particularly in critical conditions such as brain tumors. However, many healthcare institutions face challenges related to limited annotated datasets, variable image quality, and overfitting in deep learning models. This study is important because it proposes a practical and cost-effective solution that does not require collecting new data or developing complex new architectures. By optimizing image preprocessing with CLAHE and combining it with transfer learning and data augmentation, we demonstrate that diagnostic performance can be significantly improved using existing resources. Our findings contribute to more reliable and robust artificial intelligence systems in medical imaging. This is especially relevant for hospitals and research centers with limited datasets, where maximizing the value of available data is crucial. Ultimately, improving classification accuracy and model stability can enhance diagnostic confidence, reduce errors, and support safer and more efficient patient care.

Perspectives