Comparative Analysis of Transformer and CNN Architectures for OCT-Based Detection of Circular Leaf Spot Disease in Diospyros kaki

Abstract

Circular Leaf Spot (CLS) disease caused by Mycosphaerella nawae poses a significant threat to the cultivation of Diospyros kaki (persimmon). Early and accurate identification is essential to reduce yield loss, yet conventional diagnostic methods are often invasive, subjective, or time-consuming. Optical Coherence Tomography (OCT) offers a non-destructive imaging technique capable of revealing detailed leaf structures, allowing the use of deep learning models for automated disease classification. This study investigates the performance of transformer-based architectures, including ViT, Swin, CoaT, ConViT, LeViT, and MobileViT, for the classification of CLS disease OCT images. The same dataset as in previous CNN-based research was used, applying identical preprocessing techniques for fair comparison. Research was conducted using a 5-fold cross-validation, and transformer models were compared with traditional CNNs from the previous research, such as DenseNet121, VGG16, and Xception. The results show that Transformer architectures, particularly the hybrid CoaT model, achieve superior overall accuracy (87.6%) and micro-averaged AUC (0.974), surpassing the best CNN-based models. The findings highlight the strong potential of transformer architectures in the diagnosis of plant diseases using OCT, especially for early-stage detection. This research underscores the suitability of advanced vision models for high-resolution non-destructive classification tasks in agriculture.