CBMNet: a dual-attention enhanced ConvNeXt model for accurate G.V. Black type I-III classification in intraoral periapical radiographs

Abstract

Dental caries is among the most prevalent oral diseases worldwide, and accurate radiographic detection remains a clinical challenge, particularly for lesions defined by the G.V. Black classification. This study aimed to develop and evaluate CBMNet, a dual-attention enhanced ConvNeXt-Tiny model, for automated classification of G.V. Black Classes I-III using intraoral periapical radiographs. A total of 1103 anonymized periapical radiographs were retrospectively collected from the Sibar Institute of Dental Sciences, India, covering G.V. Black Class I (n = 408), Class II (n = 490), and Class III (n = 205). To address class imbalance, minority classes were supplemented with high-fidelity synthetic images generated via StyleGAN2-ADA, validated using BRISQUE scores and blinded expert review. Images were pre-processed with CLAHE and median filtering, and CBMNet was implemented by integrating Convolutional Block Attention Module (CBAM) and Multi-Scale Attention Module (MSAM) into a ConvNeXt-Tiny backbone. Hyperparameters were optimized using Particle Swarm Optimization (PSO). Performance was evaluated through stratified 5-fold cross-validation, ablation studies, and a held-out real-image test set, with additional robustness testing via test-time augmentation(TTA). CBMNet achieved a mean validation accuracy of 93.26% ( 0.81)across folds and a final held-out test accuracy of 92% with TTA. Class-wise evaluation showed high precision (Class I:0.90, Class II:0.87, Class III:0.99), recall (Class I:0.94, Class II:0.90, Class III:0.91), and F1-scores (Class I:0.92, Class II:0.89, Class III:0.95). Ablation analysis confirmed the complementary contributions of CBAM, MSAM, and TTA. Compared with baseline models (ResNet50, EfficientnetB0, DenseNet121), CBMNet consistently outperformed in overall and class-specific metrics. The proposed CBMNet framework demonstrated robust diagnostic performance for automated classification of G.V. Black Classes I-III from periapical radiographs, with accuracy and class-wise metrics exceeding 90%. By integrating dual-attention mechanisms, GAN-based augmentation, and PSO-driven optimization, CBMNet provides a reliable, interpretable, and clinically relevant tool that may support early detection and standardized diagnosis of dental caries. Future studies with multi-centre datasets and prospective clinician comparisons are warranted to further validate clinical applicability.

Keywords: Attention mechanisms; CBAM; ConvNeXt; Deep learning; Dental caries; Dental diagnostics; Intraoral periapical radiographs; MSAM; Particle swarm optimization; StyleGAN2-ADA.