Table 1 Performance of deep learning models to automate radiographic scoring in rheumatoid arthritis
From: Artificial intelligence and the future of radiographic scoring in rheumatoid arthritis: a viewpoint
Authors | Task | Training dataset | Test dataset and method | Deep learning method | Performance |
|---|---|---|---|---|---|
Morita et al. 2017 [24] Full manuscript | 1. Joint detection 2. SvdH erosion/JSN scores for MCPs and PIPs | 45 radiographs | 45 radiographs using leave-one-out cross-validation | HOG and SVM for joint detection and SVR to estimate the erosion and narrowing scores | Erosions 50.9% accuracy Absolute error 0.59 ± 0.24 JSN 64.3% accuracy Absolute error 0.43 ± 0.12 |
Morita et al. 2018 [25] Full manuscript | 1. Joint detection 2. SvdH erosion/JSN scores for MCPs and PIPs | 90 radiographs | 90 radiographs using leave-one-out cross-validation | HOG and SVM for joint detection and ridge regression to estimate the erosion and narrowing scores | Erosions 53.3% accuracy Absolute error 0.63 ± 0.32 JSN 60.8% accuracy Absolute error 0.47 ± 0/13 |
S Murakami et al. 2018 [26] Full manuscript | 1. Joint detection 2. Presence or absence of erosions | 129 radiographs | 30 radiographs, hold-out validation | MSGVF to identify regions of interest Three-layer CNN for erosion classification | Erosions Sensitivity = 0.805 Specificity = 0.9916 |
Rohrbach et al. 2019 [27] Full manuscript | 1. Ratingen erosion scores for MCPs and PIPs | 277 radiographs | 31 radiographs, hold-out validation | VGG16 inspired model | Erosions Sensitivity = 0.924 Specificity = 0.758 |
Hirano et al. 2019 [28] Full manuscript | 1. Joint detection 2. SvdH erosion/JSN scores for MCPs and PIPs | 186 training radiographs from 108 patients | 30 radiographs, hold-out validation | Uses a cascade classifier using Haar-like features to detect joints Then uses a CNN for the classification of erosions and JSN—two conv layers, two pooling, and three fully connected | Erosions Sensitivity = 0.424, 0.348 Specificity = 0.894, 0.882 JSN Sensitivity = 0.880, 0.942 Specificity = 0.748, 0.520 |
Deimel et al. 2020 [29] Abstract | 1. Joint detection 2. SvdH JSN scores for MCPs and PIPs | 5191 radiographs from 640 patients: 2207 train, 1150 validation | 1834 radiographs, hold-out validation | ROI extraction with a deep learning model that considers appearance and spatial relationship in labeling | Calculated from the confusion matrix JSN MCPs Sensitivity = 0.844 Specificity = 0.909 JSN PIPs Sensitivity = 0.863 Specificity = 0.870 |
Huang et al. 2020 [30] Abstract | 1. Joint detection 2. SvdH erosion/JSN scores for MCPs, PIPs, CMCs, and wrist | Approximately 960 hand radiographs from 309 patients diagnosed with RA | 430 radiographs from 141 patients, hold-out validation | Deep adaptive graph | JSN Sensitivity = 0.808 Specificity = 0.919 Reported explicitly, but uses cutoff joint space score ≥ 2 Data not available to calculate sens/spec for JSN vs no JSN |
Izumi et al. 2020 [31] Abstract | 1. Joint detection 2. SvdH erosion scores for PIPs, IP and MCPs | 104 x-rays | 104 radiographs, 5-fold cross-validation | CNN | 5-fold cross-validation Mean error of 0.412 per joint (of SvdH score) No further data available |