Single View Classification Results

12 views x 7 methods x 8 ML models. Click headers to sort, use search to filter.

Fusion Model Classification Results

Left/right knee fusion x 7 methods x 8 ML models.

SHAP Explainability Explorer

Select a model to view its SHAP feature importance and explanation plots.

About DisEnt-US

DisEnt-US (Disentanglement-Explainable Ultrasound Screening) is an interpretable AI framework for knee osteoarthritis assessment using ultrasound imaging. The pipeline integrates radiomics feature extraction, multi-method feature disentanglement (PCA, ICA, NMF, Factor Analysis, Sparse PCA, UMAP), and explainable machine learning to provide clinically interpretable diagnostic rules.

Key Features

• 12 standard knee ultrasound views with left/right separation
• 6 disentanglement methods for feature decomposition
• 8 machine learning classifiers with comprehensive evaluation
• Multi-view fusion models (6-view required) for enhanced diagnosis
• SHAP explainability analysis with interactive bee swarm plots
• Semantic feature mapping for clinical interpretation

Dataset

Harvard Knee Ultrasound Dataset (DOI: 10.7910/DVN/SKP9IB) — 878 subjects, 10,270 images, KL grade 0-4.

Privacy & Security

• All uploaded images are processed entirely in memory (RAM)
• No images are written to disk at any point during processing
• Uploaded data is automatically discarded after prediction completes
• Maximum upload size is limited to 16 MB per request
• No patient data is stored, logged, or transmitted to third parties
• The system operates fully offline after initial model loading

Usage Notes

• Single View: Upload one image and select the corresponding anatomical view
• Multi-View Fusion: All 6 views of the same knee side are required; upload fewer to use Single View mode instead
• Supported formats: PNG, JPG, JPEG
• For best results, use images similar to the training dataset (standard knee ultrasound views)