COMPARATIVE ANALYSIS AND EVALUATION OF THE EFFECTIVENESS OF CONVOLUTIONAL AND TRANSFORMER DEEP LEARNING ARCHITECTURES FOR MULTI-CLASS CLASSIFICATION OF MEDICAL IMAGES

Abstract

Accurate determination of the depth and area of burn damage is key to choosing the right treatment method. Doctors usually rely on visual assessment, which is often subjective and can lead to disagreements between specialists, as well as reduce the accuracy of diagnosis. Although deep learning methods for medical image analysis have been actively developed in recent years, the potential of transformer models in determining the severity of burns remains understudied. In our work, we proposed a new approach to automatic burn classification based on modern deep learning architectures. We used transfer learning technologies with pre-trained weights and applied data augmentation to improve model reliability. We chose the AdamW optimizer for training and implemented dynamic learning rate control to make the model more stable and capable of generalization. As part of our research, we tested different architectures, including convolutional neural networks and transformers. The transformer model showed the best results, achieving 94.2% accuracy, thanks to its ability to effectively take into account both local texture details and global contextual features