Table 6 Confidence-based identification of patient groups with high predictive performance for garden types III and IV in cross-validation using the internal dataset (N = 1,588). Confidence indicates the predicted probability output by deep-learning models. The Gray zone represents patient groups with confidence below 95% for both garden types III and IV. For individual networks, the best performance value for each metric is highlighted in bold. The ensemble results were obtained by combining the outputs of ResNet101, EfficientNetB4, and ResNet50, which were the top 3 models ranked based on DSC in total.
From: Garden classification of femoral neck fracture using deep-learning algorithm
Name | ACC (%) | AUC (%) | DSC (%) | Number of patients in gray zone | |||
|---|---|---|---|---|---|---|---|
Total | Confidence ≥ 95% | Total | Confidence ≥ 95% | Total | Confidence ≥ 95% | ||
EfficientNetB0 | 67.0 | 70.7 | 68.7 | 69.0 | 62.7 | 63.2 | 305 (29.3%) |
EfficientNetB2 | 67.2 | 71.7 | 71.0 | 74.5 | 63.6 | 66.4 | 279 (26.8%) |
EfficientNetB4 | 68.6 | 74.1 | 70.0 | 70.0 | 64.9 | 66.7 | 431 (41.4%) |
ResNet18 | 65.5 | 74.2 | 67.9 | 70.7 | 60.8 | 65.7 | 481 (46.2%) |
ResNet50 | 66.9 | 73.1 | 69.1 | 72.2 | 64.3 | 69.0 | 395 (37.9%) |
ResNet101 | 67.4 | 72.5 | 71.1 | 74.0 | 65.5 | 69.5 | 350 (33.6%) |
ResNext50 | 67.1 | 75.3 | 71.2 | 74.2 | 64.0 | 70.6 | 393 (37.7%) |
ReXNet100 | 63.4 | 69.0 | 65.1 | 66.4 | 60.0 | 63.1 | 336 (32.2%) |
ReXNet130 | 66.6 | 72.2 | 67.8 | 68.8 | 62.5 | 65.0 | 373 (35.8%) |
ReXNet150 | 67.9 | 75.0 | 70.4 | 71.8 | 63.9 | 68.9 | 366 (35.1%) |
DenseNet121 | 65.5 | 72.8 | 68.1 | 70.5 | 62.6 | 67.3 | 427 (41.0%) |
MobileNetV3 | 66.6 | 72.0 | 69.0 | 70.9 | 63.4 | 67.1 | 313 (30.0%) |
Ensemble of top 3 | 70.4 | 81.5 | 73.9 | 74.2 | 67.6 | 73.3 | 691 (66.3%) |
