Research map / Label-efficient and robust learning / Domain adaptation and generalization
MSSDA: Multi-Sub-Source Domain Adaptation for Diabetic Foot Neuropathy Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MSSDA, a three-stage multi-sub-source domain adaptation framework that splits source data into pseudo-domains via convolutional feature statistics and aligns selected sub-sources with the target, and releases the DFN-DS plantar pressure dataset for diabetic foot neuropathy recognition.
Problem
Cross-subject diabetic foot neuropathy recognition from plantar pressure data suffers from large domain gaps when each patient is a separate source domain, and no existing datasets contain continuous foot-specific data for DFN.
Approach
Instead of splitting sources by individual, cluster source samples with GMM over convolutional feature statistics (mean/std) from a contrastively trained encoder, pick BIC-chosen sub-sources closest to the target, and align each source-target pair with its own DANN-style sub-network.
Key results
- On DFN-DS under leave-one-subject-out CV, MSSDA reaches 0.889 accuracy and 0.921 F1, versus 0.830 for the best baseline (DAN and MCD).
- On the FRA dataset, MSSDA achieves 0.796 recall and 0.750 F1, the highest recall, with accuracy 0.729 matching MhNet's 0.730 within 0.001.
- Ablation on DFN-DS shows removing sub-source selection or multi-feature-space alignment lowers accuracy from 0.889 to 0.815-0.822.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MSSDA instead of the strongest listed baseline, DAN | DFN-DS · DFN-DS | accuracy: 0.830 -> 0.889 | Helped |
| Apply MSSDA instead of the strongest listed baseline, MhNet | FRA · FRA | F1: 0.729 -> 0.750 | Helped |
| Apply MSSDA instead of MhNet | FRA · FRA | accuracy: 0.730 -> 0.729 | Hurt |
| Add multi-feature-space alignment instead of single-feature-space alignment (ablation) | MSSDA stage 3 on DFN-DS · DFN-DS | accuracy: 0.733 -> 0.815 | Helped |
| Add sub-source domain selection to multi-feature-space alignment (ablation) | MSSDA stage 3 on DFN-DS · DFN-DS | accuracy: 0.815 -> 0.889 | Helped |
| Add sub-source domain selection to single-feature-space alignment (ablation) | MSSDA stage 3 on DFN-DS · DFN-DS | accuracy: 0.733 -> 0.822 | Helped |
| Add convolutional feature statistics to GMM clustering (ablation) | MSSDA stages 1–2 on DFN-DS · DFN-DS | accuracy: 0.770 -> 0.815 | Helped |
| Add contrastive learning to GMM clustering (ablation) | MSSDA stages 1–2 on DFN-DS · DFN-DS | accuracy: 0.770 -> 0.785 | Helped |
| Add contrastive learning when convolutional feature statistics are used (ablation) | MSSDA stages 1–2 on DFN-DS · DFN-DS | accuracy: 0.815 -> 0.889 | Helped |
| Select the two nearest sub-source domains instead of using all sub-source domains (ablation) | MSSDA on DFN-DS · DFN-DS | accuracy: 0.815 -> 0.889 | Helped |
| Select the two nearest sub-source domains rather than one nearest domain (ablation) | MSSDA on DFN-DS · DFN-DS | accuracy: 0.822 -> 0.889 | Helped |
| Select the two nearest sub-source domains rather than three nearest domains (ablation) | MSSDA on DFN-DS · DFN-DS | accuracy: 0.837 -> 0.889 | Helped |
Surprising findings
- Aligning against all sub-source domains without selection (MA alone) gave 0.815 accuracy, and selecting the wrong sub-sources (sub_1 plus sub_4) gave 0.844, indicating some source samples hurt transfer.
- MCD and MS-MDA reach 1.00 recall but 0.00 specificity on DFN-DS, collapsing to predicting all samples as DFN.
Limitations
- Accuracy drops more sharply than baselines as the decision threshold increases, which the authors attribute to excessive pursuit of transferability.
- Results on FRA are close to MhNet (accuracy 0.729 vs 0.730), and differences are small on a single dataset split.
- Number of selected sub-sources M is set by the user (M=2), with no principled selection rule.
Experimental setup
Datasets: DFN-DS Falling Risk Assessment (FRA) dataset
Benchmarks: DFN-DS FRA
Metrics: accuracy precision recall F1 specificity
Baselines: ERM D-CORAL DAN JAN MCD DANN CDAN MhNet
Builds on: DANN MFSAN M3SDA Deep CORAL CDAN
Closest papers by meaning
- Deep Hierarchies and Invariant Disease-Indicative Feature Learning for Computer Aided Diagnosis of Multiple Fundus Diseases (AAAI 2025 · similarity 0.84)
- Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data (CVPR 2025 · similarity 0.83)
- Extracting Relationships by Multi-Domain Matching (NeurIPS 2018 · similarity 0.82)
- FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.82)
- Samba: Severity-aware Recurrent Modeling for Cross-domain Medical Image Grading (NeurIPS 2024 · similarity 0.81)
- HACDR-Net: Heterogeneous-Aware Convolutional Network for Diabetic Retinopathy Multi-Lesion Segmentation (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.