Answers are grounded in this paper and cite their sources.
This study proposes a deep learning–based framework for rubric-aligned assessment of Class II cavity preparations in preclinical pediatric dentistry education. A total of 245 typodont primary first molars prepared by third-year dental students according to G.V. Black principles were scored with a structured 10-criterion rubric under standardised imaging conditions. Paired occlusal and proximal photographs were analysed using a lightweight dual-input CNN comprising a nine-criterion regression branch and a parallel three-class softmax classifier for critical-error (CE) detection (none / absence of step preparation / invalidating error). Isotonic calibration was performed on inner-validation predictions held out from CNN gradient updates. Generalisation was estimated using specimen-level nested five-fold cross-validation as the primary analysis; a fixed 80/20 partition was used only for model development and checkpoint selection. Agreement was expressed with MAE, RMSE, \(R²\) , correlation coefficients, and ICC. The contribution of the paired views was examined with a multi-seed single-view ablation, and model scores were benchmarked against a second independent examiner on the same specimens and scale. In specimen-level nested five-fold cross-validation, pooled out-of-fold predictions for all 245 specimens yielded a total-score MAE of 15.40 (95% CI 14.15–16.72), RMSE of 20.02 (95% CI 18.45–21.68), \(R^2\) of 0.244 (95% CI 0.145–0.348) and Pearson r of 0.497 (95% CI 0.392–0.601), indicating modest generalisation. The parallel critical-error classifier identified 12 of 14 invalidating errors (sensitivity 85.7%, 95% CI 57.2–98.2%; specificity 98.3%) and 28 of 33 absence-of-step errors (sensitivity 84.8%, 95% CI 68.1–94.9%) pooled across 245 out-of-fold specimens. Across the same 245 specimens, the two examiners differed by a mean absolute difference of only 1.14 points (inter-examiner \(ICC(2,1) = 0.994\) ), whereas the pooled out-of-fold model predictions deviated from the expert reference by 15.40 points—an error roughly 13 times larger. In a multi-seed ablation (means across 25 fold–seed evaluations), the dual-input configuration achieved a lower MAE (14.93) than the occlusal-only (17.71) and proximal-only (18.38) configurations and a mean \(R²\) of 0.295, whereas the single-view \(R²\) values were small. Although the pooled Pearson correlation was 0.50, this represented only a moderate association and did not establish expert-level agreement. In the 198 critical-error-free specimens, the CNN improved the subtotal MAE over constant-score baselines by approximately 1.4 points (paired bootstrap 95% CI 0.85–1.98) but could not be distinguished from a frozen MobileNetV2–Ridge baseline, indicating a modest regression-branch contribution. These findings provide preliminary evidence that rubric-related scoring patterns can be learned from paired cavity images, but not evidence of expert-level agreement. Explicit critical-error classification detected most invalidating and absence-of-step errors, yet two invalidating errors were missed—an educationally consequential limitation.
No takes yet. Share an insight, caveat, or question.
Dayan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: