Table 3 Model class-wise performance on the evaluation metric of AUROC

From: Diagnosing pathologic myopia by identifying morphologic patterns using ultra widefield images with deep learning

Model

NoPS

PS

NoMRL

TFO

DCA

PCA

MA

DeiT

96.1189 ± 0.0173

96.1254 ± 0.0172

97.7149 ± 0.0238

94.8871 ± 0.0234

94.7713 ± 0.0240

95.2838 ± 0.0433

99.3335 ± 0.0096

ConvNeXt

96.4324 ± 0.0170

96.4378 ± 0.0170

98.1800 ± 0.0227

96.0292 ± 0.0188

95.3681 ± 0.0220

96.3815 ± 0.0382

99.3870 ± 0.0102

EfficientNet

96.7455 ± 0.0158

96.7423 ± 0.0158

98.1674 ± 0.0185

95.5410 ± 0.0209

94.8309 ± 0.0247

94.1538 ± 0.0759

97.6156 ± 0.0868

Swin Transformer

95.9789 ± 0.0188

95.9643 ± 0.0189

96.7591 ± 0.0371

95.1124 ± 0.0227

94.7008 ± 0.0249

94.5249 ± 0.0592

98.5647 ± 0.0364

DINOv2

95.9202 ± 0.0172

95.9195 ± 0.0172

97.3954 ± 0.0228

95.2212 ± 0.0201

94.9398 ± 0.0228

96.3000 ± 0.0246

98.7755 ± 0.0243

VisionFM

95.6557 ± 0.0190

95.6629 ± 0.0190

96.4368 ± 0.0292

94.0472 ± 0.0254

94.2133 ± 0.0252

95.3203 ± 0.0317

98.8797 ± 0.0202

RealMNet(Ours)

97.1399 ± 0.0139

97.1671 ± 0.0139

98.3832 ± 0.0145

96.3504 ± 0.0170

95.8280 ± 0.0194

97.5641 ± 0.0199

99.1062 ± 0.0160

RealMNet-Max(Ours)

97.1369 ± 0.0140

97.1838 ± 0.0140

98.5153 ± 0.0140

96.4073 ± 0.0170

95.9304 ± 0.0186

98.0075 ± 0.0164

98.9830 ± 0.0248

  1. Reported values are the mean estimate with the standard error of the targeted metric. The estimates are computed by generating a bootstrap distribution with 1000 bootstrap samples for corresponding testing sets with n=1000 samples. The prominent results of the proposed methods are highlighted in bold.
Back to article page