fundus-9model-benchmark / analysis /ensemble_pooled.json
DoB24's picture
Pooled five-fold consistency: foundation models pooled like CNN/CLIP; 9-model McNemar (36 pairs); per-class pooled; pooled ensemble report
767b053 verified
Raw
History Blame Contribute Delete
588 Bytes
{
"protocol": "pooled five-fold soft vote, weight = pooled macro-F1, all nine models",
"members": [
"inception_v3",
"clip_openai",
"vgg19",
"resnet101",
"dinov2_l",
"densenet121",
"resnet50",
"swin_b",
"retfound"
],
"ensemble": {
"acc": 90.43,
"f1": 92.71,
"roc_auc": 0.9945,
"ece": 0.0237,
"kappa": 0.887
},
"best_single": {
"model": "inception_v3",
"acc": 90.18,
"f1": 92.54
},
"ensemble_vs_best_mcnemar": {
"b": 54,
"c": 46,
"p": 0.484,
"significant_005": false
},
"n_test": 3208
}