Table 1 Summary of reviewed literature with a focus on dataset split and reported test classification performance.
From: Inflation of test accuracy due to data leakage in deep learning-based classification of OCT images
Ref. | OCT dataset | Data split strategy | Model performance on testing set |
|---|---|---|---|
Thyroid, parathyroid, fat and muscle samples | per-image | 97.12% accuracy | |
Pituitary adenoma | per-image | 0.96 AUC | |
Ophthalmology15* (version 2) | original split | 95.55% accuracy | |
Ophthalmology15* (version 2) | original split | 99.1% accuracy | |
Ophthalmology15* (version 3) | original split | 98.7% accuracy | |
Ophthalmology15* (version 3) | original split | 96.6% accuracy | |
Ophthalmology15* (train version 2, test version 3) | original split | 99.6% accuracy | |
(1) Ophthalmology15* (train version 2, test version 3) (2) Ophthalmology16* | (1) original split (2) per-volume/subject | (1) 99.80% accuracy (2) 100% accuracy | |
Coronary artery | per-volume/subject | 96.05% accuracy | |
Kidney† | per-volume/subject | 82.6% accuracy | |
High and low grade brain tumors | per-volume/subject | 97% accuracy | |
Colon**† | per-volume/subject | 88.95% accuracy on 2D images | |
Breast tissue | per-volume/subject | 91.7% specificity | |
Ophthalmology15* (version 2) | per-volume/subject | 98.46% accuracy | |
(1) Ophthalmology15* (version 3) (2) Ophthalmology16* (3) Breast tissue14* | (1) original split (2) per-volume/subject (3) per-image | (1) ~96% accuracy (2) >98.8% accuracy (3) 98.8% accuracy | |
(1) Ophthalmology16* (2) Ophthalmology41* (3) Ophthalmology42* (4) Ophthalmology15* (unclear version) | (1) per-volume/subject (2) per-volume/subject (3) per-volume/subject (4) original split | (1) 96.66% accuracy (2) 98.97% accuracy (3) 99.74% accuracy (4) 99.78% accuracy | |
Dentistry | No description given | 98% sensitivity 100% specificity | |
Ophthalmology | No description given | 99.19% accuracy |
