A spatial adaptive multi-scale ConvNeXt framework for robust, calibrated, and statistically validated multi-class retinal OCT classification

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Accurate classification of retinal diseases from optical coherence tomography (OCT) images is important for clinical decision support. However, many studies in deep learning put the primary focus on enhancing accuracy and often provide limited attention to calibration and robustness under real-world variations. This work presents an enhanced ConvNeXt-Tiny framework incorporating a Spatial Adaptive Multi-scale Convolution (SAMC) module for retinal OCT classification. Rather than introducing a new convolutional operation, SAMC combines established multi-scale dilated convolution and feature fusion principles in a lightweight stage-wise configuration within the ConvNeXt hierarchy to refine representations of retinal abnormalities occurring at different spatial scales. The model is evaluated on the OCT-C8 dataset with eight retinal disease classes. It obtains a test accuracy of 98.39% and a macro F1 score of 0.9839. Reliability of classification was also analyzed by measuring Expected Calibration Error (ECE) and Brier score. Temperature scaling is applied as a post-hoc calibration method, while Monte Carlo dropout is used to estimate predictive variability under stochastic inference. Robustness is evaluated under noise, blur, and illumination changes, where minimal performance drop was observed. Computational complexity analysis is performed in terms of model parameters, floating-point operations, and inference efficiency to assess the practicality of the proposed framework. Statistical significance is confirmed using McNemar's test. Interpretability analysis with Grad-CAM and LIME reveals the model attends to clinically significant regions of the retina. Overall, the proposed ConvNeXt-SAMC framework achieves accurate retinal disease classification, while additional calibration, uncertainty, robustness, statistical, and interpretability analyses provide a broader assessment of model behavior.