Artificial Intelligence-Based Medical Image Classification Using Convolutional Neural Networks

Abstract

Artificial intelligence has become an important computational approach for supporting image-based diagnosis because convolutional neural networks (CNNs) can learn hierarchical visual features directly from radiological, ophthalmic, dermoscopic and pathological images. This methodology paper proposes a reproducible CNN-based framework for multiclass medical image classification using transfer learning. The procedure combines image quality screening, resizing, intensity normalization, class-balanced augmentation, patient-level data partitioning, transfer learning with pretrained ResNet50 and DenseNet121 backbones, and evaluation using accuracy, precision, recall, F1-score, sensitivity, specificity and area under the receiver operating characteristic curve (AUC). The methodology is designed for a public labelled image dataset such as HAM10000 or a disease-specific X-ray/OCT collection, while emphasizing prevention of data leakage and clinically meaningful evaluation. An illustrative pilot comparison shows how a transfer-learned CNN can improve discrimination relative to a small CNN trained from scratch, while also demonstrating the importance of class-wise sensitivity and AUC rather than accuracy alone. The study concludes that CNN-based classification offers strong potential as a decision-support technology, but external validation, explainability, calibration and clinician oversight remain essential before clinical use.