DL predictis mandibular growth using cephalometric radiographs

1. Title:

Deep learning-based prediction of mandibular growth trend in children with anterior crossbite using cephalometric radiographs

2. Authors: Jia-Nan Zhang, Hai-Ping Lu, Jia Hou, Qiong Wang Feng-Yang Yu, Chong Zhong, Cheng-Yi Huang and Si Chen

3. Summary:

The study aimed to develop a deep learning CNN model to automatically predict whether the mandible of a child with anterior crossbite would grow to be normal or overdeveloped using cephalometric radiographs. The model, built on the ResNet50 algorithm, was trained on 256 radiographs and tested on 40. It achieved an accuracy of 85%, significantly outperforming junior orthodontists, who averaged 54.2%. Visualization via Grad-CAM indicated that the model primarily analyzes the chin, lower mandibular edge, incisor teeth, airway, and condyle to make its predictions

4. Key Words:

Deep learning; Mandibular growth; Prediction; Anterior crossbite; Cephalometric radiographs

5. Extracted data

5.1. Year: 2023

5.2. Modality: Cephalometric radiography

5.3. Dataset: 296 patients

5.4. Dataset Split: 256 training / 40 testing

5.5. Network Architecture: ResNet50

5.6. Metrics: Metrics: Accuracy; Sensitivity (TPR); Specificity (TNR); AUC

5.7. AP - Professional Qty: 5 orthodontists (2 experts for tracing + 3 juniors for classification)

5.8. AP - Supervisor Presence: Yes

5.9. AP - Experience Level: Not specified

5.10. AP - Expertise Area: Orthodontist

5.11. AP - Tool or System: Dolphin Imaging Software

5.12. ML Task: Classification / Prediction

5.13 Project Objective: To develop a deep learning model to automatically predict mandibular growth outcomes to assist in clinical treatment planning and prognosis for children with anterior crossbite

6. Clinical Relevance: Predicting mandibular growth is critical because the mandible has the longest growth period of the craniofacial bones. Accurate early prediction helps orthodontists decide between interventional orthodontic treatment and orthognathic surgery. Inaccurate predictions by humans can lead to unsuccessful "camouflage" treatments that may relapse or deteriorate; the deep learning model provides a more accurate, objective tool to reduce these clinical risks