+ HEALTH Aims Intramuscular fat (IMF) is increasingly recognized as a marker of ectopic adiposity and adverse cardiome…+ HEALTH Periapical lesions are challenging to detect on intraoral radiographs because of anatomical superimposition a…± HEALTH Ultrasound imaging is an indispensable diagnostic tool, yet its profound reliance on operator expertise inher…+ HEALTH Objectives We adapted the individualized polysocial risk score (iPsRS), a machine learning model originally d…± HEALTH Melioidosis is a life-threatening infectious disease caused by Burkholderia pseudomallei ( Bp ). Rapid diagno…+ HEALTH Classification is a core machine learning task, and ensemble voting methods are widely used to improve predic…+ CLIMATE Microplastic (MP) pollution poses escalating environmental risks, demanding efficient and reproducible tools…± HEALTH Medicinal chemistry sits at the center of modern drug discovery, yet translating molecular designs into appro…
TruaceTracing the truth around AIWednesday, August 26, 2026
Subscribe
Health·G Space·Evidence-backed gain·Published 2026-08-26

Stacked Ensemble Deep Learning Models for Accurate Detection and Size Stratification of Periapical Lesions on Intraoral Radiographs

Abstract: Periapical lesions are challenging to detect on intraoral radiographs because of anatomical superimposition and reader variability. This study developed stacked deep learning ensembles for automated detection and radiographic size stratification of periapical lesions. In total, 146 radiographs comprising normal cases and three lesion-size categories were cropped around the root apex and augmented using predefined transformations. Five convolutional neural network backbones were trained, and their probability out…

TRV-2026-0899Peer-reviewedPermanent record — cite & verify
Stacked Ensemble Deep Learning Models for Accurate Detection and Size Stratification of Periapical Lesions on Intraoral Radiographs

"Intraoral Periapical Radiograph (IOPA) showing Deciduous(Milky or Primary) Tooth 75 and developing crown of Permanent or Secondary Teeth 35, 36 and 37" by Nizil Shah is licensed under CC BY-SA 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/4.0/.

The quick read

Researchers built stacked deep learning ensembles to detect periapical lesions and stratify them by radiographic size on intraoral radiographs. Using 146 cropped and augmented images, five CNN backbones were combined with MLR and XGBoost meta-learners and evaluated on an internal hold-out test set.

High internal accuracy and sensitivity for sub-2 mm lesions suggests potential value for radiographic decision support in endodontics, but the small single-centre design means results are preliminary. By the August 2026 publication date, no prospective multicentre external validation or clinical deployment outcome had been reported.

Main points
  • Study used 146 radiographs covering normal cases and three lesion-size categories, cropped around the root apex and augmented with predefined transformations.
  • Five convolutional neural network backbones were trained and combined with MLR and XGBoost meta-learners.
  • Internal hold-out testing reported MLR accuracy 0.96 and macro-averaged ROC-AUC 1.00, XGBoost accuracy 0.95 with similar AUC.
Gain

Stacked ensembles combining five CNN backbones improved automated detection and size stratification of periapical lesions on cropped intraoral radiographs, reaching high accuracy and high sensitivity for very small lesions on internal testing.

The rundown

The authors curated 146 intraoral radiographs and cropped them around the root apex, applying predefined augmentation. Five CNN backbones were trained to classify normal versus three lesion-size categories, then stacked using multinomial logistic regression and gradient boosting meta-learners.

On the internal hold-out test set both ensembles outperformed individual backbones, with MLR reporting 0.96 accuracy and 1.00 macro-averaged ROC-AUC and XGBoost 0.95 accuracy. Sensitivity for lesions under 2 mm was 0.94, addressing a known difficulty from anatomical superimposition and reader variability.

Sources

Reader signal

How should this claim be treated?

The debate