Hierarchical Deep Decision Tree-Based Network for Odontogenic Cystic Lesion Classification in CBCT Images

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Odontogenic cystic lesions (OCLs) are complex jaw abnormalities requiring accurate diagnosis for effective treatment.
• Current diagnosis relies on cone-beam computed tomography (CBCT) analysis of morpho-pathological features, often overlooking hierarchical relationships.
• Existing AI methods lack guidance and do not fully utilize the hierarchical nature of OCL diagnosis.

Purpose of the Study:

• To propose a novel hierarchical deep decision tree network (H2DT-Net) for improved OCL classification.
• To leverage inter-categorical relationships and integrate diagnostic and morpho-pathological features.
• To enhance feature extraction through lesion-focused attention maps.

Main Methods:

• Developed H2DT-Net with three modules: Deep Hierarchical Learning Module (DHLM), Feature Category Embedding Module (FCEM), and Lesion Localised Attention Module (LLAM).
• DHLM leverages inter-categorical relationships for hierarchical learning.
• FCEM captures representations from diagnostic and morpho-pathological domains; LLAM generates lesion-focused attention maps.

Main Results:

• H2DT-Net achieved state-of-the-art performance in OCL classification on 289 CBCT images.
• The model demonstrated superior diagnostic accuracy compared to six maxillofacial clinicians in a clinical setting.
• The proposed modules effectively captured hierarchical relationships and focused feature extraction.

Conclusions:

• H2DT-Net represents a significant advancement in AI-assisted OCL diagnosis.
• The hierarchical approach and attention mechanism improve classification accuracy.
• H2DT-Net shows strong potential for clinical application in maxillofacial diagnostics.