Table 2 We applied different performance measures to the attention-based MIL classifier. They show a clear benefit from using Kaggle challenge data for pre-training. ROC AUC and PR AUC are the area under the receiver operating characteristic and the precision recall curve, respectively. Searching for the optimal input resolution of the images (values between \(299 \times 299\) px and \(1299 \times 1299\) px) shows that \(799 \times 799\) px produces the best results. Highest ROC AUC over models is highlighted in bold.
Performance measure | ||||||
|---|---|---|---|---|---|---|
Precision | Recall | Accuracy | F1 | ROC AUC | PR AUC | |
Model with ImageNet Pre-Training (Full resolution) | 0.788 | 0.772 | 0.674 | 0.78 | 0.653 | 0.859 |
Model with KAGGLE Pre-Training (Full resolution) | 0.880 | 0.802 | 0.770 | 0.839 | 0.810 | 0.925 |
Full Model with down-sampled data (\(299 \times 299\) px, ablation study) | 0.809 | 0.713 | 0.659 | 0.758 | 0.628 | 0.840 |
Full Model (\(799 \times 799\) px) | 0.954 | 0.822 | 0.837 | 0.883 | 0.890 | 0.964 |
