Automatic 3D landmarking using deep neural networks on CT scans

TITLE:

Automatic 3D landmarking model using patch-based deep neural networks for CT image of oral and maxillofacial surgery

Authors: Qingchuan Ma; Etsuko Kobayashi; Bowen Fan; Keiichi Nakagawa; Ichiro Sakuma; Ken Masamune; Hideyuki Suenaga

SUMMARY:

This study proposes a patch-based deep neural network for fully automatic 3D cephalometric landmark detection in CT images of oral and maxillofacial surgery (OMS) patients. The authors argue that manual landmarking remains highly time-consuming and dependent on specialist expertise, particularly in OMS cases where anatomical deformities are common. A dataset of 66 CT volumes was prepared, including segmentation of head skeletons in MIMICS and manual annotation of 13 key anatomical landmarks. Landmark coordinates were standardized and reduced using PCA to incorporate anatomical correlations into the training process. The model employs a 2.5D patch-based representation to reduce computational burden while retaining spatial information, using a CNN with joint regression and classification subnets to iteratively refine landmark positions. Trained on 58 volumes and tested on 8, the network achieved an average landmarking error of 5.785 mm and a mean inference time of 37.871 seconds per CT volume. Performance varied by anatomical condition: symmetric structures produced near-expert accuracy, whereas severe deformities or scanning noise led to larger deviations. The authors highlight that, although errors remain above the ideal threshold for certain surgical tasks, the model's performance is aligned with previous literature and demonstrates clinical feasibility—particularly for applications that tolerate deviations of up to 2–6 mm. This work provides a foundation for future OMS automation pipelines and has potential to reduce surgeon workload and improve workflow efficiency.

KEY WORDS:

machine learning; 3D cephalometry; landmark detection; CNN; OMS

EXTRACTED DATA:

Year: 2020 Modality: CT
Dataset: 66 CTs
**Dataset Split:**50 train / 8 val / 8 test Architecture: CNN (3-layer backbone + regression & classification subnets)
**Metrics:**Average Accuracy (mm), processing time (s)
AP Qty: 1 Supervisor: No information Expertise: OMF surgeon
Tool/System: MIMICS
ML Task: Object detection (3D landmarking) Objective: Automate OMS cephalometric landmarking to reduce workload and improve consistency..

CLINICAL RELEVANCE:

Potential to reduce surgeon workload and improve consistency in OMS planning.