Dual-model deep learning for detecting and classifying dental abscesses, cysts, and tumours on orthopantomograms: a histopathology-referenced diagnostic study

Differential diagnosis of periapical and intra-osseous jaw lesions is a recurring clinical challenge that still depends on invasive, costly histopathological examination. Deep learning applied to routine panoramic radiography offers a potential non-invasive triage aid, but most prior work has addressed a single lesion class or a single algorithmic task. To evaluate, on orthopantomograms, a two-task artificial-intelligence (AI) framework intended as a first-reader triage and decision-support aid that (i) classifies a lesion as abscess, cyst, or tumour and (ii) independently localises and segments lesions, and to characterise how class imbalance constrains each task. This work extends a previously published single-class (abscess) analysis by the same group to a three-class, dual-model setting. In a retrospective, multi-centre diagnostic study, panoramic radiographs with osseous lesions were curated and cross-referenced with histopathology reports to establish the reference standard. A classification model (EfficientNet-B3, transfer learning) was trained on the histopathologically labelled set and expanded by a three-level augmentation strategy. A separate localisation/segmentation model (YOLOv8-Seg) was trained on manually annotated bounding boxes and polygon masks. The two models were trained and evaluated independently on a held-out test set ( n = 90: 40 abscesses, 40 cysts, 10 tumours); reporting follows the CLAIM and CLAIRE recommendations for AI in medical imaging. During training the augmented EfficientNet-B3 reached 96.40% accuracy on the training data, but on the independent test set overall accuracy fell to 71.1% and balanced accuracy to 53.3%, underscoring that the training-phase figure is optimistic. Per-class recall was 87.5% (95% CI 73.9–94.5) for abscesses and 72.5% (57.2–83.9) for cysts, but the classifier recalled no tumour case (0%; 0–27.8). YOLOv8-Seg localised lesions (detection efficiency) at 65.0% (abscess), 37.5% (cyst), and 60.0% (tumour), but correctly typed them (sensitivity) at only 50.0%, 32.5%, and 30.0%, respectively. Both architectures handled the well-represented abscess class adequately and performed poorly on the scarce tumour class. Task-specialised, independently trained models are a workable strategy for radiographic lesion analysis, but performance is bounded by the size and balance of the training data. The concordant weakness of two structurally different models on the tumour class is consistent with data scarcity and class imbalance being a major limiting factor, although an independent architectural contribution cannot be excluded from these data. The findings argue for larger, balanced, multi-institutional datasets before clinical deployment.