Conditional ordinal regression based on chain rule probability in Swin Transformer for improved ordinal image classification

• Effective image analysis relies on robust feature extraction, a capability for which Swin Transformer is suited. However, ordinal image classification specifically requires capturing the ordered progression of labels. Conventional multiclass methods ignore the natural order of labels and treat them as independent categories, treating all misclassifications equally and leading to severe prediction errors. • To overcome this limitation, Swin Transformer with Conditional Ordinal Regression (CORN) is proposed to model the ordered progression of categories. This framework decomposes the ordinal task into a sequence of conditional binary classifications using chain rule probability. This enables the model to learn the dependency relation between categories and reduces the prediction errors. The proposed Swin+CORN framework is examined on three medical image datasets, namely RetinaMNIST, Alzheimer and Knee Osteoarthritis, and compared with standard Swin-based multiclass classification and Swin+CORAL as an alternative ordinal regression approach. • Swin+CORN achieves the best overall performance over all the three datasets. The proposed framework consistently outperforms both Swin+Multiclass and Swin+CORAL in terms of accuracy and prediction error. Experimental results show the efficiency of the hierarchical visual representation learning combined with CORN for the ordinal image categorisation task.