CrossSpine: Multi-scale Cross-sequence Attention with Anatomical Priors for Automated Pfirrmann Grading

Technical Summary: CrossSpine

Problem Statement
Automated grading of Lumbar Disc Degeneration (LDD) is critical for the objective quantification of structural changes associated with low back pain. While Magnetic Resonance Imaging (MRI) is the clinical standard, utilizing the Pfirrmann grading system, existing deep learning baselines have demonstrated insufficient performance on specific datasets. The authors identify two primary limitations in current approaches:

1. Single-stream processing: Standard models fail to effectively integrate complementary features from different MRI sequences (e.g., T1, T2, etc.), treating them in isolation rather than leveraging their mutual information.
2. Lack of anatomical context: Baseline models often treat all lumbar discs uniformly, ignoring the biomechanical and degeneration patterns that are distinct to specific intervertebral disc (IVD) levels.

Methodology
The authors propose CrossSpine, a framework designed to address these limitations through multi-scale cross-sequence attention and anatomical integration. The architecture consists of four main stages:

1. Data Preprocessing: The system utilizes a curated dataset of 1,185 IVD records from 237 patients. For each examination, the middle image of four available MRI sequences is selected and resized to 224×224 pixels. Data augmentation techniques (flipping, rotation, scaling, etc.) are applied to improve generalization.
2. Multi-Scale Feature Extraction: Using a DenseNet-121 backbone pretrained on ImageNet, the model extracts features at two distinct spatial scales:
   • Scale 1: Features extracted after the first dense block (56×56 spatial resolution), reduced to a 128-dimensional vector.
   • Scale 2: Features extracted after the second dense block (28×28 spatial resolution), reduced to a 256-dimensional vector.
3. Cross-Sequence Attention Fusion: Instead of simple concatenation, the framework employs a cross-sequence attention mechanism (inspired by Vaswani et al.) to adaptively fuse features from the four MRI sequences. This allows sequences to reference one another, enabling the network to learn which sequences are most diagnostically informative for specific features. This process is applied independently at both spatial scales.
4. Scale Fusion and IVD-Aware Classification:
   • The feature vectors from both scales are projected to a common 512-dimensional space and combined using learnable scale weights.
   • IVD-Aware Classification: A novel technique explicitly integrates anatomical disc-level information (e.g., L1–L2, L5–S1) into the classification head. The fused features are projected to a 10-dimensional representation, concatenated with the disc-level indicator, and passed through a final linear layer to predict the Pfirrmann grade. This allows the model to learn level-specific degeneration priors.
   • The system handles missing sequences via zero-padding and availability masks to prevent information leakage.

Key Contributions

1. CrossSpine Architecture: A novel framework utilizing cross-sequence attention to fuse multi-scale features from different MRI sequences, enhancing the capture of subtle degenerative changes.
2. Curated Dataset: The introduction of a meticulously curated dataset (PhenMec) specifically designed for automated Pfirrmann grading, featuring balanced anatomical representation across disc levels despite clinical class imbalances.
3. IVD-Aware Classification: A model-agnostic technique that injects anatomical disc-level information into the classification process. The authors demonstrate that this improves performance not only in their proposed architecture but also in baseline models.

Experimental Results
Experiments were conducted on an NVIDIA RTX 4090 GPU using a 70/15/15 train/validation/test split. The evaluation prioritized Macro F1-score, Mean AUROC, and Mean AUPRC to account for class imbalance.

• Performance Gains: Compared to a DenseNet-121 baseline (without IVD awareness or cross-sequence attention), CrossSpine achieved:
  • A 125% relative improvement in Macro F1 score (0.1917 to 0.4318).
  • A 99% increase in Mean AUPRC (0.2772 to 0.5516).
  • A 36% increase in Mean AUROC (0.6187 to 0.8391).
• Ablation Studies: The study confirmed that both the Multi-scale architecture and the IVD-Aware module are critical. Adding IVD awareness alone nearly doubled the baseline Macro F1, while combining both modules yielded the highest performance.
• Backbone Analysis: Among tested backbones (ResNet-34, EfficientNet-B0, DenseNet-121), DenseNet-121 provided the best consistency and performance within the proposed framework.
• Qualitative Analysis: Visualizations showed that while baseline models frequently misclassified discs by assigning the majority class (Grade 2), CrossSpine successfully captured level-specific patterns, correctly identifying intermediate disc grades where baselines failed.

Significance
The paper claims that CrossSpine establishes a robust benchmark for multi-modal medical imaging by demonstrating that integrating anatomical context (disc level) and multi-sequence attention significantly enhances diagnostic performance. The work underscores the necessity of moving beyond single-stream, uniform treatment of spinal data to leverage the rich, complementary information inherent in multi-sequence MRI and the biomechanical realities of specific spinal levels. The IVD-aware technique is highlighted as a standalone, model-agnostic contribution that can be applied to other spine imaging architectures.