Learning curves in transfer-learning dental radiograph classification with pre-trained and random features

Introduction: The number of annotated images required to train a dental radiograph classifier is rarely quantified, and it is unknown whether transfer learning changes how fast performance saturates. Methods: Pre-trained and randomly initialized ResNet-18 features were extracted from 182 annotated teeth in the DENTEX dataset, and a logistic regression classifier was trained on stratified subsamples of 5, 10, 20, 40, 80 and 127 teeth over five seeds, with test accuracy recorded at each size. Results: With pre-trained features, accuracy rose from 50.2% at 5 teeth to 67.3% at 127 teeth, reaching 95% of full-data accuracy at 80 teeth. With random features, accuracy reached 58.2% at 127 teeth and plateaued earlier, at 40 teeth, but never matched the pre-trained curve. At 20 training teeth, pre-trained features classified caries at 80.7% but deep caries at 6.0% and no periapical lesion correctly. Conclusions: Transfer learning with pre-trained features reaches its plateau only around 80 annotated teeth, later than a random-feature baseline, and the small-sample performance is concentrated in the majority class, so the number of annotations a dental task requires depends on the class balance rather than on the overall accuracy alone.