Abstract
BACKGROUND: Accurate medical image segmentation significantly impacts patient outcomes, especially in diseases such as skin cancer, intestinal polyps, and brain tumors. While deep learning methods have shown promise, their performance often varies across datasets and modalities. Combining advanced segmentation techniques with traditional feature extraction approaches may enhance robustness and generalizability. OBJECTIVE: This study aims to develop an integrated framework combining segmentation, advanced feature extraction, and transfer learning to enhance segmentation accuracy across diverse medical imaging (MI) datasets, thus improving classification accuracy and generalization capabilities. METHODS: We employed independently trained U-Net models to segment skin cancer, polyps, and brain tumor regions from three separate MI datasets (HAM10000, Kvasir-SEG, and Figshare Brain Tumor dataset). Moreover, the study applied classical texture-based feature extraction methods, namely Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrix (GLCM), processing each Red Green Blue (RGB) channel separately using an offset [0 1] and recombining them to create comprehensive texture descriptors. These segmented images and extracted features were subsequently fine-tuned pre-trained transfer learning models. We also assessed the combined performance on an integrated dataset comprising all three modalities. Classification was performed using Support Vector Machines (SVM), and results were evaluated based on accuracy, recall (sensitivity), specificity, and the F-measure, alongside bias-variance analysis for model generalization capability. RESULTS: U-Net segmentation achieved high accuracy across datasets, with particularly notable results for polyps (98.00%) and brain tumors (99.66%). LBP consistently showed superior performance, especially in skin cancer and polyp datasets, achieving up to 98.80% accuracy. Transfer learning improved segmentation accuracy and generalizability, particularly evident in skin cancer (85.39%) and brain tumor (99.13%) datasets. When datasets were combined, the proposed methods achieved high generalization capability, with the U-Net model achieving 95.20% accuracy. After segmenting the lesion regions using U-Net, LBP features were extracted and classified using an SVM model, achieving 99.22% classification accuracy on the combined dataset (skin, polyp, and brain). CONCLUSION: Integrating deep learning-based segmentation (U-Net), classical feature extraction techniques (GLCM and LBP), and transfer learning significantly enhanced the accuracy and generalization capabilities across multiple MI datasets. The methodology provides robust, versatile framework applicable to various MI tasks, supporting advancements in diagnostic precision and clinical decision-making.