Deep learning-based automatic scoring models for the disease activity of rheumatoid arthritis based on multimodal ultrasound images
Authors' conclusion
DL models based on multimodal US images allow a quantitative and objective assessment of RA activity.
Study context
- Population
- Multimodal ultrasonographic (US) images of RA scored by EULAR-OMERACT Synovitis Scoring system1,244 training images
- Comparison
- Deep learning (DL) models based on static greyscale (SGS), dynamic greyscale (DGS), static power Doppler (SPD) and dynamic power Doppler (DPD) US images vs radiologists, scored against EULAR-OMERACT Synovitis Scoring by two expert radiologists
- Outcome
- Best-performing model per synovitis score 0/1/2/3 (DPD, SGS, DGS, SPD respectively), AUC
Methodological detailsdeep learning (ResNet-type) models, two separate test cohorts, comparison with 12 radiologists
- Deep learning (ResNet-type) models
- Two separate test cohorts
- Comparison with 12 radiologists
Performance
As reported in the abstract
| Test or model | AUC (score 0) | AUC (score 1) | AUC (score 2) | AUC (score 3) |
|---|---|---|---|---|
| DPD | 0.87 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract |
| SGS | — not stated in the abstract | 0.95 | — not stated in the abstract | — not stated in the abstract |
| DGS | — not stated in the abstract | — not stated in the abstract | 0.74 | — not stated in the abstract |
| SPD | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract | 0.95 |
AUC (score 0)
DPD0.87
SGS— not stated in the abstract
DGS— not stated in the abstract
SPD— not stated in the abstract
AUC (score 1)
DPD— not stated in the abstract
SGS0.95
DGS— not stated in the abstract
SPD— not stated in the abstract
AUC (score 2)
DPD— not stated in the abstract
SGS— not stated in the abstract
DGS0.74
SPD— not stated in the abstract
AUC (score 3)
DPD— not stated in the abstract
SGS— not stated in the abstract
DGS— not stated in the abstract
SPD0.95
— not stated in the abstract
Further findings
| Outcome | SGS | SPD |
|---|---|---|
| SPD vs SGS, test cohort 1, AUC for scores 0/2/3 | ||
| AUC (score 0) | 0.66 | 0.82 |
| AUC (score 2) | 0.66 | 0.67 |
| AUC (score 3) | 0.75 | 0.95 |
| SPD vs SGS, test cohort 2, AUC for score 0 | ||
| AUC (score 0) | 0.81 | 0.89 |
- SPD vs SGS, test cohort 1, AUC for scores 0/2/3: test cohort 1 (152 images)
- SPD vs SGS, test cohort 2, AUC for score 0: test cohort 2 (354 images)
Agreement of DL models with experienced radiologists, per image (intraclass correlation coefficient)
- All DL models comparable to experienced radiologists; intraclass correlation coefficient
- 0.239-0.756
- P < 0.05
SPD vs SGS, test cohort 1, AUC for scores 0/2/3
- SPD model performed better than SGS
Dynamic vs static DL models and vs senior radiologists
- Dynamic DL models performed better than static ones in most scoring processes and were more accurate than most senior radiologists, especially the DPD model
Original publication
Deep learning-based automatic scoring models for the disease activity of rheumatoid arthritis based on multimodal ultrasound images. Rheumatology (Oxford). Published 20 July 2023. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →