AI detection of radicular cysts and granulomas on panoramic X-rays

A Deep Learning Approach for Radiological Detection and Classification of Radicular Cysts and Periapical Granulomas

Authors: Jonas Ver Berne, Soroush Baseri Saadi, Constantinus Politis, Reinhilde Jacobs

---

Summary This study from KU Leuven and Karolinska Institutet presents a two-route deep learning framework for detecting and differentiating radicular cysts and periapical granulomas on panoramic radiographs.
Differentiating these two common apical lesions is clinically critical: cysts require surgical enucleation, while granulomas are usually treated by root canal therapy.

The dataset comprised 152 confirmed lesions (80 cysts / 72 granulomas) plus 255 control images (197 normal + 58 other radiolucent lesions).
Each panoramic image was cropped into:

1. Global image – half of the mandible containing the lesion.
2. Local image – close-up view of the lesion itself.

The AI framework combined:

• MobileNetV2 (classification network) – 53 layers, dual input (global + local).
• YOLOv3 (localization network) – 53-layer Darknet architecture for bounding-box detection.

Training used 249 panoramic radiographs, split into 90 % train / 10 % test, with augmentation (optical distortion, tone curve, noise, flip, etc.) and 10-fold cross-validation.

Results:

• Radicular cysts: Sensitivity = 1.00 (95 % CI 0.63–1.00); Specificity = 0.95 (0.86–0.99); AUC = 0.97
• Periapical granulomas: Sensitivity = 0.77 (0.46–0.95); Specificity = 1.00 (0.93–1.00); AUC = 0.88
• Localization (YOLOv3): Average Precision (AP) = 0.83 for cysts and 0.74 for granulomas
• Hardware: Intel i9-12900K CPU, 128 GB RAM, NVIDIA Quadro P5000 GPU (16 GB)

The two-route network achieved reliable clinical performance, combining both global context and local detail.
Grad-CAM visualization confirmed that the CNN focused on both lesion borders and internal features.

DOI: https://doi.org/10.1016/j.jdent.2023.104581

Key Words

Panoramic Radiography, Radicular Cyst, Periapical Granuloma, Deep Learning, MobileNetV2, YOLOv3, Lesion Localization, Computer Vision, Dental AI, Artificial Intelligence in Dentistry

Extracted Data

##- Year: 2023
##- Modality: Panoramic Radiography
##- Dataset: 249 images
##- Dataset Split: 90 % training | 10 % test (10-fold CV)
##- Network Architecture: MobileNetV2 (53 layers) + YOLOv3 (53 Darknet layers)
##- Metrics: AUC; Sensitivit; Specificity
##- AP – Strategy: Consensus annotation by 3 OMF radiologists
##- AP – Professional Qty: 3 experts (2 annotators + 1 arbiter)
##- AP – Supervisor Presence: Yes
##- AP – Experience Level: Experienced OMFR specialists
##- AP – Expertise Area: Dentomaxillofacial Radiology
##- AP – Tool or System: No information
##- Task: Lesion Detection + Classification (two-route multiscale DL framework)
##- Project Objective: To develop and validate a deep learning workflow for radiographic differentiation of radicular cysts and periapical granulomas on panoramic imaging.

---

Clinical Relevance

##- Clinical importance: Cysts and granulomas require different treatments (surgery vs endodontic therapy). Visual distinction on panoramic X-rays is challenging.
##- Innovation: A two-route deep learning system combining global and local features achieves near-expert diagnostic accuracy and adds localization via YOLOv3.

• Practical impact: Automated radiological support can enhance referral efficiency and standardize diagnostic decisions in general and specialized dental practice.

DOI: https://youtu.be/6GonbTjFS2A