Sclerosis Classification Methods

• Sclerosis classification is a set of computational and imaging techniques to detect and subtype tissue hardening in conditions like MS, ALS, and glomerulosclerosis.
• Advanced methods utilize deep learning, transformer architectures, and multimodal fusion to achieve high accuracy, with metrics such as AUC ~0.834 and >80% accuracy in some cases.
• Ongoing research focuses on domain adaptation, data imbalance mitigation, and improved interpretability to enhance diagnostic precision and clinical prognostication.

Sclerosis classification encompasses a diverse array of computational, imaging, and molecular methods for the detection, subtyping, and severity staging of sclerotic processes across neurological, musculoskeletal, and renal pathologies. These methods address the classification of demyelinating lesions in multiple sclerosis (MS), neurodegenerative hallmarks in amyotrophic lateral sclerosis (ALS), renal glomerular injury in glomerulosclerosis, and fibrotic signatures in systemic sclerosis and related multisystem diseases. Recent advances draw upon multimodal data integration, ML pipelines, and rigorous statistical modeling to improve diagnostic accuracy, subphenotype resolution, and clinical prognostication.

1. Clinical Contexts and Sclerosis Phenotypes

Sclerosis, operationally defined as a pathological hardening or scarring of tissue, arises in multiple organ systems with distinct clinical implications:

• Neurological Disorders:
  • Multiple Sclerosis (MS): Characterized by multi-focal demyelinating lesions in the CNS white matter, leading to staging via the Expanded Disability Status Scale (EDSS) and functional subscores (FS). Disease subtypes include relapsing-remitting (RRMS), secondary-progressive (SPMS), primary-progressive (PPMS), radiologically isolated syndrome (RIS), and clinically isolated syndrome (CIS) (Vazifehdan et al., 22 Jan 2025, Mato-Abad et al., 2024, Costa et al., 2020).
  • Amyotrophic Lateral Sclerosis (ALS): Progressive loss of upper and lower motor neurons, with brain MRI increasingly investigated for structural classifiers distinguishing ALS from controls (Kushol et al., 2023, Elahi et al., 2017).
• Renal Pathology:
  • Global Glomerulosclerosis (GGS): Global obliteration and fibrosis of glomeruli; subtypes include obsolescent, solidified, and disappearing sclerosis, with direct impact on chronic kidney disease prognosis (Lu et al., 2022, Lu et al., 2021).
• Systemic Autoimmune/Fibrotic Disease:
  • Systemic Sclerosis (SSc): Multisystem fibrotic vasculopathy, with classification driven by molecular and optical markers (França et al., 2019).

The detection, subtype discrimination, and longitudinal tracking of these entities constitute the core tasks of sclerosis classification.

2. Imaging-Based Classification Methodologies

Magnetic resonance imaging (MRI) and computed tomography (CT) are principal tools for lesion characterization and detection:

• MS Lesion Classification with Bayesian Spatial Models: Binary lesion maps derived from $T_2$-weighted MRI are analyzed using spatially varying coefficient probit models. This approach incorporates subject covariates (MS subtype, age, gender, disease duration, EDSS) and an MCAR prior over spatial coefficients $\{\beta(s_j)\}$, fitted via Gibbs sampling and data augmentation. Five-way MS subtype classification accuracy achieved $0.772\pm0.052$ overall, with $0.828\pm0.047$ average per-subtype accuracy—substantially outperforming mass-univariate logistic models and naive Bayes (Ge et al., 2014).
• ALS Classification with Transformer Architectures: The SF$^2$Former pipeline integrates a vision transformer (ViT) branch (spatial features) and a GFNet branch (frequency-domain MRI features), fusing 2D coronal slice embeddings for subject-level majority-voting classification. On multi-center CALSNIC MRI datasets, accuracy rates for ALS vs. control reach $81.6\%$–$81.8\%$, with F1-scores exceeding baselines (ResNet, CoHOG). Frequency-augmented fusion and careful plane selection are critical for identifying subtle neurodegenerative changes (Kushol et al., 2023).
• Automated Sclerotic Metastasis Detection in CT: Coarse-to-fine cascades begin with highly sensitive candidate generators utilizing rule-based segmentation and SVM committees, followed by convolutional neural networks (CNNs) that aggregate $N=100$ randomly-perturbed 2D views per candidate ROI, averaging network outputs to robustly filter false positives. The two-stage approach yields area-under-the-ROC-curve (AUC) of $0.834$ and reduces false positives by a factor of 3–4 at $60$–$\{\beta(s_j)\}$0 sensitivity (Roth et al., 2014).
• Renal Glomerulosclerosis Subtype Classification: Detection uses CircleNet for glomerulus localization in whole-slide images (WSI), followed by fine-grained classification with ResNet-V2 models pretrained at scale (BiT-M, ImageNet-21k). Class-imbalance is addressed through mini-batch oversampling and focal-loss. Macro-F1 reaches $\{\beta(s_j)\}$1, with AUC of $\{\beta(s_j)\}$2 on external binary validation (Lu et al., 2022).

3. Machine Learning Pipelines for Clinical Staging and Severity

Sclerosis severity staging often leverages clinical, imaging, and laboratory data through ML pipelines:

• Multiple Sclerosis (EDSS) Staging and Progression Prediction:
  • Longitudinal Imputation and SVM Regression: Functional system subscores (FS) are frequently missing in real-world MS registries. Both exponential weighted moving average (EWMA, RMSE $\{\beta(s_j)\}$3) and classification and regression trees (CART, RMSE $\{\beta(s_j)\}$4) are benchmarked for imputation. Post-imputation, radial basis function (RBF) SVMs trained on all features achieve $\{\beta(s_j)\}$5, $\{\beta(s_j)\}$6 (CART+SVM pairing superior), enabling precise patient-stage stratification (Vazifehdan et al., 22 Jan 2025).
  • Multimodal Deep Neural Networks: Integrating MRI (five sequences), structured EHR, and free-text clinical notes, modality-specific encoders (3D-ResNet for MRI, self-attention CNN for EHR, GNN for notes) feed into a bi-GRU fusion-decoder. Five-fold cross-validation yields AUROC $\{\beta(s_j)\}$7 for EDSS$\{\beta(s_j)\}$8 milestone classification—up to a $\{\beta(s_j)\}$9 gain over single-modality models. Notes and FLAIR MRI contribute the majority of predictive signal (Zhang et al., 2023).
• Text-Based Severity Classification: The MS-BERT domain-specific transformer, further trained on $0.772\pm0.052$0 de-identified MS consult notes, underpins the MSBC classifier, aggregating chunkwise [CLS] embeddings via a CNN-Seq2Vec module. Macro-F1 for EDSS classification reaches $0.772\pm0.052$1, and functional subscore macro-F1 is $0.772\pm0.052$2, outperforming rule-based and Word2Vec-CNN baselines (Costa et al., 2020).
• Patient-Reported Outcome Measures (PROMs): An elastic net classifier selects a sparse subset of PROMs (16/145), with multi-output regression predicting future PROMs. Prognostic accuracy for RR-SP course transition is about $0.772\pm0.052$3, suggesting PROM-driven models can anticipate MS progression well ahead of standard quantitative exams (Fiorini et al., 2016).

4. Molecular and Spectral Biomarker-Based Sclerosis Classification

High-dimensional molecular and spectroscopic data drive new avenues in sclerosis subtyping:

• Transcriptomic Classifiers for MS: XGBoost models trained on bulk and single-cell transcriptomics (CD4$0.772\pm0.052$4, B-cells, PBMC, CSF) following rigorous normalization, batch correction (ComBat, SCGen), and gene-declustering (correlated features $0.772\pm0.052$5) yield AUC up to $0.772\pm0.052$6 (B-cells, CSF). Feature importance via SHAP uniquely identifies immune checkpoints (ITK, CLEC2D, KLRG1, CEACAM1), translation/ribosomal programs, and lipid metabolism nodes overlapping with EBV pathways. SHAP priorities are complementary to differential expression methods, and the overlap with known etiopathogenic axes supports mechanistic relevance (Massafra et al., 5 Mar 2026).
• Optical Spectral Classification in Systemic Sclerosis: Diffuse reflectance near-infrared spectroscopy (NIRS) across $0.772\pm0.052$7–$0.772\pm0.052$8 nm is processed via moving-average filtering, baseline correction, and normalization. Recursive feature elimination (RFECV) with linear SVC consistently identifies $0.772\pm0.052$9 nm as the discriminative wavelength (singlet oxygen luminescence; ^1O_2). Median accuracy for Proximal Interphalangeal Joints is $0.828\pm0.047$0 across repeated train/test splits (França et al., 2019).

5. Advanced Methodologies for Sclerosis Subtype and Tissue-Level Classification

Novel computational and learning strategies enhance subtype resolution and tissue specificity:

• Fine-Grained Glomerulosclerosis Subtyping:
  • CircleMix Data Augmentation: Effective for highly imbalanced class distributions, this augmentation mixes mini-batch images via randomly generated circular sector masks, with mixing ratio proportional to the sector's angular width. Hierarchical EfficientNet-B0 classifiers trained with CircleMix augmentation achieve five-class balanced accuracy of $0.828\pm0.047$1, surpassing CutMix and non-hierarchical baselines (Lu et al., 2021).
  • Holistic Detectors: The combination of detection (CircleNet) and classification (ResNet101-BiT-M) supports simultaneous call-out, subtype assignment, and false-positive filtering, validated on large WSI datasets and robust to stain/population variation (Lu et al., 2022).
• MRI Microstructure Modeling for Tissue Lesion Detection: Multimodal voxelwise classifiers integrate Spherical Mean Technique (SMT) for intra-/extra-axonal diffusion MRI compartment estimation, myelin water imaging via multi-exponential T2 relaxometry, and robust AdaBoost tree ensembles. Lesion vs normal-appearing white matter accuracies reach $0.828\pm0.047$2, with intra-neurite fraction and myelin water fraction as dominant discriminants. The approach is validated on manual expert segmentations and supports automated lesion-probability mapping (Fischi-Gomez et al., 2021).

6. Current Limitations and Future Directions

Despite significant advances, several limitations persist:

• Generalizability: Multicenter and cross-platform imaging data require harmonization protocols; transformer and ensemble models can exhibit performance drops ($0.828\pm0.047$3) when domain shift is not addressed (Kushol et al., 2023).
• Sample Sizes and Imbalance: Most state-of-the-art pipelines are developed on restricted or enriched cohorts. Transfer learning and data augmentation (e.g., CircleMix) mitigate class imbalance but require further external validation (Lu et al., 2021, Lu et al., 2022).
• Interpretability: While SHAP and attention mechanisms provide insights into model predictions and feature importance, full mechanistic interpretability—in particular, identifying true pathophysiological drivers rather than associations—remains an open research area (Massafra et al., 5 Mar 2026, Zhang et al., 2023).
• Temporal and Prognostic Modeling: Survival-oriented, event-time models are needed for predicting progression or transition between sclerosis subtypes on a patient level (Fiorini et al., 2016, Zhang et al., 2023).
• Modality Integration: Cross-modality fusion (e.g., MRI, EHR, clinical notes, omic) offers the highest discriminatory performance but increases system complexity; effective attention and gating mechanisms for modality weighting are ongoing targets for methodological innovation (Zhang et al., 2023, Kushol et al., 2023).

A plausible implication is that further gains in sclerosis classification will depend on scalable multimodal data acquisition, robust domain adaptation, interpretable model architectures, and systematic validation across diverse populations and pathologies.