Image-Level Classification of Central Vein Categories on FLAIR MRI in Multiple Sclerosis Using a Lightweight Explainable Deep Learning Framework

This is the authors' abstract. We don't add key points for this paper.

Background/Objectives: The central vein sign can help characterize multiple sclerosis (MS) lesions on MRI, but reading it takes specialist time and experience. We developed a compact convolutional neural network (CNN) and a deep feature engineering (DFE) pipeline for three-class classification of clinically labeled MR images. Methods: The MS Vein dataset contains 3181 axial FLAIR images from 154 participants in vein-positive, vein-negative, and control classes. MACDIBNeXt (about 6.5 million parameters) combines pooling-based attention downsampling, grouped convolution, and bottleneck and inverted bottleneck blocks. It was pretrained on Stable ImageNet1k and fine-tuned on a participant-level split (98 training, 25 validation, and 31 test participants). The DFE pipeline extracted 768 global average pooling features, selected 47 with CWINCA in participant-grouped folds of the training data, and classified them with tkNN. Results: On 635 held-out test images from 31 participants, the standalone CNN classified 527 correctly (82.99%; participant-cluster bootstrap 95% CI, 77.93–87.66%). DFE classified 571 correctly (89.92%; 85.47–93.82%), with 89.13% macro F1 and Cohen’s κ of 0.83. With participants resampled as clusters, the 6.93-point gain had a 95% CI of 3.33–10.63 points (p = 0.0002), and a cluster-adjusted McNemar test gave p = 0.0021. Grad-CAM activation was intracranial in the displayed examples. Conclusions: MACDIBNeXt with DFE gives promising internal, image-level results. Because the data are retrospective, single-center, and labeled on FLAIR images, external participant-level validation is required before clinical use.