Optimizing Skin Lesion Diagnosis Using Deep Learning and Novel Noise Image Filtering Algorithm

Abstract

Non-Reference Image Quality Assessment (NR-IQA) has made a significant contribution to medical imaging ervices such as X-ray, Magnetic Resonance Imaging (MRI), and Computed Tomography (CT) scans. Integrating deep learning artificial intelligence algorithms into NR-IQA has revolutionized medical services, enhanced diagnostic accuracy, and improved patient care management. Among deep learning algorithms, Convolutional Neural Networks (CNNs) are extensively utilized for object detection, image recognition and classification, and semantic segmentation. However, the CNN algorithm has a crucial challenge with noisy datasets that can decrease decision-making accuracy. Thus, applying NR-IQA to CNN algorithms can enhance feature extraction and diagnosis accuracy. This research proposes an efficient skin lesion diagnosis by integrating NF-IQA into the CNN algorithm. The proposed system calculates NR-IQA using three optimal metrics: average information entropy, chromatic level factor, and average luminance. By filtering low-quality images, NR-IQA reduces the noisy images in both the training and testing datasets, which enables the CNN to focus on clinically relevant features. Additionally, it is designed to diagnose seven types of skin lesions. The proposed system performance is evaluated using Area Under the Curve-Receiver Operating Characteristic (AUC-ROC), confusion matrix, and classification report, which have been used to measure accuracy, loss, precision, recall, and F1 score. The proposed system outperforms the baseline methods and achieves accuracies of 100%, 93.3%, and 85% on balanced training, validation, and testing datasets, respectively. The main advantage of the proposed system is to improve the learning efficiency and robustness of CNN-based classifiers, and to offer rapid skin lesion diagnosis, thereby enhancing AI-driven medical care.