Treffer: Enhancing lumbar disc herniation classification through region-of-interest guidance and geometric shape features.

Lumbar disc herniation (LDH) is one of the most common degenerative diseases of the spine. Magnetic resonance image is the most effective way to detect LDH. The variety of shapes and blurred boundaries of diseased discs, along with the unclear classification basis of existing methods and their poor ability to differentiate between lesion types, make computer-aided diagnosis (CAD) of LDH challenging. We propose an enhanced classification of LDH through region-of-interest guidance and geometric shape features (RGGS-Net) to address these challenges. RGCG-Net establishes the connection between the segmentation of diseased lumbar disc and the classification of lesion types in LDH. A region-of-interest guided module, combined with region-of-interest supervision, is proposed to refine the features from the encoder. Weighted skip connections are used to balance the ratio between the original feature and the refined feature. Hierarchical supervision is used to reduce the training difficulty of the deep decoder and improve the final segmentation performance. Finally, the precise classification of LDH is achieved based on the geometrical features of its different types. Numerous experiments have demonstrated the effectiveness of the RGGS-Net. The classification accuracy of the RGGS-Net in the LDH classification task is 0.965. The Dice of the RGGS-Net reaches 0.957 in vertebrae and disc segmentation task.
(© 2026 IOP Publishing Ltd. All rights, including for text and data mining, AI training, and similar technologies, are reserved.)