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

INTRODUCTION

Annotated data is the principal bottleneck in dental artificial intelligence, because every lesion, tooth and restoration in a training set must be labelled by a clinician whose time is scarce, and because the open dental collections that do exist are far smaller than the natural-image corpora on which general-purpose models are trained.(1,2) A research group that wants to build a caries detector or an implant classifier must therefore decide how many radiographs to annotate before the marginal benefit of another label stops justifying its cost, and in practice that decision is usually made by convention, habit or available budget rather than by measurement.(3)

The cost of that decision is asymmetric: annotating too little produces a model that fails silently on the classes it has barely seen, while annotating too much wastes scarce clinician effort that could have been spent on the next task, so a measured answer to the sample-size question is a practical necessity rather than a purely methodological nicety. The question of how many annotated images are needed is a learning-curve question, and it has a concrete answer for a given dataset and pipeline, yet it is seldom answered in the dental imaging literature.( 4)

Transfer learning changes the shape of that curve. Instead of training a network from scratch, the standard dental pipeline reuses a backbone pre-trained on ImageNet and trains only a lightweight head, so that the model begins with a representation of visual structure that no amount of dental data could teach it from a small sample.(5,6)

The pre-trained backbone provides generic features, edges, textures and shapes, that transfer to a radiographic target even though ImageNet contains no radiographs, because the low-level statistics of natural images overlap sufficiently with those of medical images.(7-9) Studies of transfer learning in medical imaging have shown that pre-trained features often outperform features trained from scratch when data are scarce, and that the advantage depends on the similarity between the source and target domains, with natural-image features transferring well to radiographs despite the visual gap. What is less clear is how fast the transfer-learning advantage saturates, and whether a dental classifier reaches its ceiling after a handful of examples or only after hundreds, which is the distinction that determines whether annotation effort should be concentrated or spread thin.(4,10)

The question is sharpened by the class structure of dental diagnoses, because a curve that plateaus on overall accuracy may still be climbing on the rare classes that matter clinically, and the sample size at which the overall curve flattens is not necessarily the sample size at which every clinically important class has been adequately learned.

The present study contrasted two feature representations, namely pre-trained ResNet-18 features and randomly initialized ResNet-18 features, across a range of training sizes, so that the contribution of pre-training could be separated from the contribution of the architecture itself.

This contrast is informative because the two representations share exactly the same dimensionality and the same downstream classifier, and differ only in whether the backbone weights carry the statistics of natural images, which cleanly isolates the effect that transfer learning is claimed to provide.(9,10)

A random-feature baseline is a stronger control than no baseline at all, because it matches the architecture and therefore distinguishes the benefit of pre-training from the benefit of simply having a fixed nonlinear feature map, a distinction that is lost when transfer learning is compared only against a hypothetical from-scratch network that is never actually trained. The study also traced the curve separately for each diagnosis class, because a single overall curve can conceal the divergent behaviour of a majority class that saturates early and a rare class that remains unlearned, and the per-class view was expected to be at least as informative as the overall one. The objective was to trace the learning curve of a dental radiograph classifier as a function of the number of annotated training teeth and to identify the sample size at which performance plateaus, comparing pre-trained and random features.

METHODS

RESULTS

Pre-trained versus random features across training sizes

The pre-trained representation outperformed the random representation at almost every one of the six training sizes tested, and the gap between the two curves behaved differently at small and large sample sizes, widening as the sample grew rather than shrinking as the data accumulated (Table 2).

Training teeth	Pre-trained features (%)	Random features (%)
5	50.2	43.3
10	53.8	44.0
20	56.4	50.2
40	56.4	59.3
80	64.7	58.5
127	67.3	58.2

Note: Accuracy is mean across five stratified subsamples. The full training split contains 127 teeth. Source: Elaborated from the learning-curve analysis (DENTEX).

With 5 training teeth the pre-trained features reached 50.2% accuracy compared with 43.3% for the random features, a difference of 6.9 percentage points. By 127 training teeth the pre-trained features reached 67.3% compared with 58.2% for the random features, so the gap widened to 9.1 percentage points at the largest size. The pre-trained advantage was therefore present from the very smallest sample and did not narrow as data accumulated over the six training sizes, which indicates that the benefit of pre-training is not a small-sample crutch that disappears once the target data become adequate but a persistent property of the representation.

The two curves also differed in their trajectories across the six training sizes (Figure 1). The random-feature curve rose to a peak of 59.3% at 40 training teeth and then declined slightly to 58.2% at the full 127 teeth, whereas the pre-trained curve was flat between 20 and 40 teeth at 56.4% and then rose again to 64.7% at 80 teeth and 67.3% at the full sample.

The pre-trained curve therefore showed a two-phase pattern, with an early plateau followed by a later gain, while the random-feature curve rose to its peak and then reversed direction after 40 training teeth, declining slightly at the two largest sizes. The two curves crossed only in the middle of the range: at 40 teeth the random representation briefly matched the pre-trained representation, at 59.3% against 56.4%, before the pre-trained curve pulled ahead again at the larger sizes, so the ranking between the two representations was not monotonic across the full range of training sizes.

The subsample-to-subsample variability was largest at the very smallest training sizes, reaching a standard deviation of 12.9 percentage points for the random representation at 5 teeth, and it narrowed as the sample grew, which is the expected consequence of a small training set whose composition varies substantially from draw to draw.

DISCUSSION

One would expect a transfer-learning classifier to saturate within a handful of examples, on the reasoning that its backbone is already trained and only a small head needs to adapt to the dental domain; the learning curve instead showed that the pre-trained representation did not reach 95% of its full-data accuracy until 80 annotated teeth, and that below 20 teeth the classifier was effectively a caries detector that failed on the other three classes.

The random-feature baseline, by contrast, reached its own plateau earlier and at a lower ceiling, which clarifies what pre-training actually contributes to the learning curve: not an earlier plateau, but a higher and later one that continues to improve as more dental examples arrive.

This reversal of the intuitive expectation is the central contribution of the study, because it means that the convenience of transfer learning should not be mistaken for an exemption from the data-collection problem, and that a research group relying on a pre-trained backbone still faces the unavoidable task of accumulating enough annotated examples to cover the rare classes it cares about before the model can be considered clinically complete.

The contrast between the pre-trained and random representations is the clearest summary of the results. Compared with the random features, the pre-trained features began with a 6.9 percentage-point advantage at 5 teeth and widened that advantage to 9.1 points at 127 teeth, which is consistent with the established finding that ImageNet pre-training provides a useful starting representation for medical tasks but does not by itself teach the domain-specific features that a dental classifier needs.(6,7)

The widening gap also matches the observation that the benefit of a better or better-matched source model grows with the target dataset size, so that pre-training and annotation effort are complements rather than substitutes.(9,10) In the random-feature group the curve peaked at 40 teeth and then reversed, a sign that the random representation began to overfit the noise in the larger training sample rather than extracting additional signal, whereas the pre-trained curve, in the same training sizes, continued to rise.

The reversal of the random-feature curve is itself informative: it shows that a non-pre-trained representation can be made worse by more data when the representation is too noisy to benefit from it, a failure mode that the pre-trained curve did not exhibit. This non-monotonic behaviour is a reminder that a learning curve need not be increasing, and that an observed decline with more data is a diagnostic signal that the representation, rather than the amount of data, is the limiting factor, which is exactly the situation that transfer learning is designed to remedy.

The two-phase shape of the pre-trained curve, flat between 20 and 40 teeth and then rising again after 40, is not an artifact but a consequence of the class imbalance documented in the dataset.

The first phase, in which accuracy reached about 56%, is the point at which the classifier has learned the majority class, caries, which alone accounts for more than half the teeth; the second phase, beginning after 40 teeth, is where enough minority-class examples have accumulated for the classifier to begin distinguishing impacted teeth and deep caries, which is why the overall accuracy did not reach 95% of its ceiling until 80 teeth.

This reading is supported directly by the timing: 20 teeth contain roughly 11 caries examples but only about 1 or 2 periapical examples, so a classifier trained on 20 teeth has effectively seen the majority class and little else, and a single periapical example is not enough for a linear classifier to learn a reliable boundary for that class. The per-class results at 20 teeth make this reading concrete, because the 80.7% accuracy on caries coexisted with 6.0% on deep caries and 0.0% on periapical lesions, so the early plateau was a majority-class plateau rather than a genuine saturation of the task.(3,4) The ordering of the per-class accuracies followed the class prevalence almost exactly, with the most prevalent class the most accurate and the rarest class the least accurate, which is the signature of a classifier whose performance is governed by the number of examples per class rather than by any intrinsic diagnostic difficulty of the classes themselves.

The comparison with the wider medical imaging literature places these results firmly in context. Transfer learning has been reported to require fewer examples than training from scratch for tasks as varied as pneumonia, osteoporosis and kidney classification, and the compendious reviews of transfer learning for cancer diagnosis consistently find a pre-training advantage, but those studies rarely report the full learning curve or a plateau size, so the present result that pre-training delays rather than advances the plateau is a refinement rather than a contradiction.( 11-14)

The same pattern recurs in studies that compare transfer-learning pipelines for osteoporosis prediction and in ensemble transfer approaches for Alzheimer diagnosis, which report the accuracy of the final model without tracing how that accuracy accumulates with data.(15,16) Work on effective data selection and on multi-domain meta-datasets has begun to ask the complementary question of which examples matter most, and it suggests that the value of an added label is not uniform, which is consistent with the majority-class concentration observed here and with the idea that active or targeted annotation of the rare classes would bend the curve upward faster than random annotation.(17,18)

The present results therefore connect the sample-size question to the data-selection question: if each added label for a rare class is worth more than a label for the majority class, then the optimal annotation strategy is to oversample the rare classes, and the learning curve provides the evidence for that strategy by showing how little the majority-class labels contribute once the majority class has already saturated.

The present curve therefore sits within an emerging literature that treats both the number and the choice of training examples as first-class design variables rather than as an afterthought of model selection or an unexamined default.

The contrast with the classical sample-size literature is instructive: where a conventional power calculation asks how many subjects are needed to detect an effect at a given significance, a learning-curve analysis asks how many examples are needed to reach a given accuracy, and the two questions give different kinds of answers, the first probabilistic and the second empirical.(3) The learning curve is in this sense the more direct instrument for a builder of classifiers, because it reports the observed relationship between data and performance rather than an assumption about that relationship, and it can be re-estimated whenever the dataset, the set of classes or the training pipeline changes.(4)

The dental-specific literature likewise reports accuracy without reporting how that accuracy depends on the number of annotated teeth, so the present learning curve provides the sample-size evidence that the field has been missing.( 19-21)

Dental studies of ankylosis, supernumerary teeth and other uncommon findings are particularly exposed to the sample-size problem, because their target classes are rare and their reported accuracy may rest on a small number of positive examples.(22,23) Comparative evaluations of vision transformers and convolutional networks for dental radiographs similarly focus on architecture rather than on the amount of data, so the present curve complements them by isolating the data dimension.( 24)

The finding that a small annotated sample buys majority-class performance while leaving the rare classes essentially unlearned is directly relevant to the many dental studies that train on a few hundred images and report a single overall accuracy, because that single number can conceal the same majority-class concentration observed here.( 25,26)

The practical consequence for annotation planning is that the number of images a dental task requires cannot be read from a single target accuracy without knowing the class balance. A group that needs only to flag caries, the majority class, may find that a few dozen labelled teeth suffice, because the majority class saturates early, whereas a group that needs to detect periapical lesions or deep caries must annotate enough images to accumulate rare-class examples, which in this dataset meant approaching the full sample before the overall accuracy approached its ceiling. This distinction between majority-class and rare-class sample requirements is the main actionable message of the curve, and it is the reason the plateau size alone, without the per-class breakdown, would be misleading.(3,4)

The message is amplified by the shape of the curve rather than by any single point on it: a group that stops annotating when the overall accuracy first flattens, at around 20 to 40 teeth, will believe it has reached the ceiling when it has in fact only finished learning the majority class, and will leave the clinically important rare conditions undetected. A more reliable rule of thumb is to monitor the accuracy on the rarest class that the model must detect and to continue annotating until that class-specific curve, not the overall curve, reaches its own plateau, even though that threshold will typically demand many more labelled examples than the overall accuracy would suggest.

The strengths of this study are the paired design, in which the pre-trained and random features were extracted from the same teeth and passed through the same split, scaler and classifier, and the use of an openly licensed dataset that makes the curve reproducible from the public source alone. The principal limitation is the size of the dataset, because 182 teeth and a 55-tooth test set make the per-class accuracies at small training sizes noisy, and a larger multi-institution collection would trace the minority-class curves with useful precision rather than leaving the periapical class with a zero at the small sizes.

A second limitation is that the analysis used a single backbone and a single downstream classifier, so the exact plateau size of 80 teeth should not be transplanted to a different architecture or a different dental task without re-measuring it; the qualitative two-phase pattern, however, is likely to generalize because it follows from class imbalance rather than from the specifics of ResNet-18.(1,5) A third limitation is that the random-feature baseline used the same architecture as the pre-trained backbone, and a comparison against a genuinely task-trained-from-scratch network or a foundation model would place the transfer-learning advantage on a fuller scale.(27,28)

A fourth limitation is that the learning curve was measured with a single fixed test split, so the absolute accuracy values carry the variance of that particular split, even though the comparison between the two feature representations is protected by the paired design. These limitations do not change the central finding, but they bound how far the specific numbers should be generalized, and they point toward a larger, multi-class, multi-institution learning-curve study as the natural extension.

The implication for dental artificial intelligence is that annotation effort should be guided by the class that is hardest to learn rather than by the class that is easiest, and that a learning curve reported with a per-class breakdown is a more honest guide to data requirements than a single accuracy at a single size. Where rare conditions are the clinical target, the curve suggests that annotation should continue well past the point at which overall accuracy appears to plateau, because the overall plateau masks the continued learning of the minority classes, and stopping at that point would leave those classes effectively undetected.

This recommendation holds broadly across health systems, because the class imbalance that drives it is a property of disease prevalence rather than of any particular institution, and a small dental service in a low-resource setting faces the same majority-class concentration when it trains on a modest labelled set.(29,30)

The portability of the recommendation follows from the fact that the two-phase curve arises from prevalence, which is broadly similar across populations for the common dental conditions studied here, even though the absolute class sizes and the achievable accuracies will differ between settings. The learning curve itself is a portable tool rather than a local one: any dental group can trace its own curve with the same subsampling procedure on its own labelled collection, and the resulting plateau and per-class breakdown will be specific to that group's prevalence and imaging protocol, which is precisely the information that a generic sample-size recommendation cannot provide.

CONCLUSIONS

A transfer-learning dental radiograph classifier with pre-trained ResNet-18 features reached 95% of its full-data accuracy only at 80 annotated teeth, whereas a random-feature baseline plateaued earlier, at 40 teeth, but at a lower accuracy, so pre-training delayed the plateau while raising its ceiling.

The pre-trained representation outperformed the random representation at every training size, with an advantage that widened from 6.9 to 9.1 percentage points as the sample grew, which shows that pre-training and annotation effort are complements rather than substitutes.

At small training sizes the accuracy was concentrated in the majority class, caries, while the minority classes remained largely unlearned, so the number of annotations a dental task requires is governed by the class balance and by the rarest condition that must be detected rather than by the overall accuracy alone, and a learning curve with a per-class breakdown is the appropriate instrument for planning that annotation effort.

References

1. 1 Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, et al. A survey on deep learning in medical image analysis. Medical Image Analysis. 2017;42:60-88. doi: 10.1016/j.media.2017.07.005
2. 2 Ibrahimi Hamamci et al. DENTEX: an abnormal tooth detection with dental enumeration and diagnosis benchmark for panoramic X-rays [Internet]. 2023. Available from: https://huggingface.co/datasets/ibrahimhamamci/DENTEX
3. 3 Figueroa RL, Zeng-Treitler Q, Kandula S, Ngo LH. Predicting sample size required for classification performance. BMC Med Inform Decis Mak. 2012;12(1):8. doi: 10.1186/1472-6947-12-8
4. 4 Viering T, Loog M. The shape of learning curves: a review [Internet]. 2023. Available from: https://arxiv.org/abs/2103.10948
5. 5 He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016:770-778. doi: 10.1109/cvpr.2016.90
6. 6 Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in deep neural networks? [Internet]. 2014. Available from: https://arxiv.org/abs/1411.1792
7. 7 Tajbakhsh N, Shin JY, Gurudu SR, Hurst RT, Kendall CB, Gotway MB, et al. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?. IEEE Trans Med Imaging. 2016;35(5):1299-1312. doi: 10.1109/tmi.2016.2535302
8. 8 Shin HC, Roth HR, Gao M, Lu L, Xu Z, Nogues I, et al. Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Trans Med Imaging. 2016;35(5):1285-1298. doi: 10.1109/tmi.2016.2528162
9. 9 Raghu M, Zhang C, Kleinberg J, Bengio S. Transfusion: understanding transfer learning for medical imaging [Internet]. 2019. Available from: https://arxiv.org/abs/1902.07208
10. 10 Kornblith S, Shlens J, Le QV. Do better ImageNet models transfer better? [Internet]. 2019. Available from: https://arxiv.org/abs/1805.08974
11. 11 Katreddi S, Midatani A, Roy AP, Velpuri U, Kasani S. Pediatric pneumonia X-ray image classification: predictive model development with DenseNet-169 transfer learning. J Med Artif Intell. 2025;8:37-37. doi: 10.21037/jmai-24-356
12. 12 Qureshi MB, Sani M, Raza A, Qureshi MS, Alghamdi W, Akbar S, et al. Deep‑learning based osteoporosis classification in knee X‑rays using transfer‑learning approach. Sci Rep. 2025;15(1):38448. doi: 10.1038/s41598-025-24338-4
13. 13 Pimpalkar A, Saini DKJB, Shelke N, Balodi A, Rapate G, Tolani M. Fine-tuned deep learning models for early detection and classification of kidney conditions in CT imaging. Sci Rep. 2025;15(1):10741. doi: 10.1038/s41598-025-94905-2
14. 14 Kaur N, Hans R. Transfer Learning for Cancer Diagnosis in Medical Images: A Compendious Study. Int J Comput Intell Syst. 2025;18(1):62. doi: 10.1007/s44196-025-00772-0
15. 15 Kaur N, Khan S, Alazman I, Bin-Asfour M, Alam MN, Balasubramaniam V, et al. Transfer learning based osteoporosis prediction using enhanced medical imaging and fuzzy fusion. Sci Rep. 2025;15(1):41960. doi: 10.1038/s41598-025-25946-w
16. 16 Varzaneh ZA, Mousavi SM, Khoshkangini R, Moosavi Khaliji SM. An ensemble model based on transfer learning for the early detection of Alzheimer’s disease. Sci Rep. 2025;15(1):34634. doi: 10.1038/s41598-025-22025-y
17. 17 Lee H, Kwak JY, Lee E. Effective data selection via deep learning processes and corresponding learning strategies in ultrasound image classification. Sci Rep. 2025;15(1):16138. doi: 10.1038/s41598-025-00416-5
18. 18 Woerner S, Jaques A, Baumgartner CF. A comprehensive and easy-to-use multi-domain multi-task medical imaging meta-dataset. Sci Data. 2025;12(1):666. doi: 10.1038/s41597-025-04866-4
19. 19 Raeisi Z, Rokhva S, Rahmani F, Goodarzi A, Najafzadeh H. Multi-label diagnosis of dental conditions from panoramic x-rays using attention-enhanced deep learning. Oral Maxillofac Surg. 2025;29(1):166. doi: 10.1007/s10006-025-01463-y
20. 20 Pornprasertsuk-Damrongsri S, Vachmanus S, Papasratorn D, Kitisubkanchana J, Chaikantha S, Arayasantiparb R, et al. Clinical application of deep learning for enhanced multistage caries detection in panoramic radiographs. Sci Rep. 2025;15(1):33491. doi: 10.1038/s41598-025-16591-4
21. 21 Eftimie S, Ileni T, Iacob L, Hedeșiu M, Dioșan L. Tooth-level detection and mapping of dental pathologies on panoramic radiographs using YOLOv11 and RT-DETR. MethodsX. 2025;15:103696. doi: 10.1016/j.mex.2025.103696
22. 22 N. Y. Artificial intelligence for detecting dental ankylosis in primary molars using panoramic radiographs—a retrospective study. JOCPD. 2025;49(6):120. doi: 10.22514/jocpd.2025.133
23. 23 Uzel İ, Ghabchi B, Çoğulu D. Deep learning-based automated detection of supernumerary teeth in pediatric panoramic radiographs. PLoS One. 2025;20(11):e0335845. doi: 10.1371/journal.pone.0335845
24. 24 Felek T, Tercanlı H, Gök RŞ. Evaluating vision transformers and convolutional neural networks in the context of dental image processing: a systematic review. BMC Oral Health. 2025;25(1):1626. doi: 10.1186/s12903-025-07036-5
25. 25 Sukegawa S, Yoshii K, Hara T, Tanaka F, Taki Y, Inoue Y, et al. Optimizing dental implant identification using deep learning leveraging artificial data. Sci Rep. 2025;15(1):3724. doi: 10.1038/s41598-025-87579-3
26. 26 Pishghadam N, Esmaeilyfard R, Paknahad M. Explainable deep learning for age and gender estimation in dental CBCT scans using attention mechanisms and multi task learning. Sci Rep. 2025;15(1):18070. doi: 10.1038/s41598-025-03305-z
27. 27 Tavakoli N, Shakeri Z, Gowda V, Samsel K, Bedayat A, Ghasemiesfe A, et al. Generative AI and Foundation Models in Radiology: Applications, Opportunities, and Potential Challenges. Radiology. 2025;317(2):e242961. doi: 10.1148/radiol.242961
28. 28 Wang Z, Sun Q, Zhang B, Wang P, Zhang J, Zhang Q. PM2: A new prompting multi-modal model paradigm for few-shot medical image classification. Computer Methods and Programs in Biomedicine. 2025;272:109042. doi: 10.1016/j.cmpb.2025.109042
29. 29 Benakatti V, Nayakar RP, Anandhalli M, Sukhasare RC. Efficacy of deep learning models and dental professionals in identifying dental implants. Imaging Sci Dent. 2025;55(4):351. doi: 10.5624/isd.20250048
30. 30 Achararit P, Manaspon C, Jongwannasiri C, Kulthanaamondhita P, Itthichaisri C, Chantarangsu S, et al. Impacted lower third molar classification and difficulty index assessment: comparisons among dental students, general practitioners and deep learning model assistance. BMC Oral Health. 2025;25(1):152. doi: 10.1186/s12903-025-05425-4