Project: Improved automated segmentation of dental CBCT images with Auto3DSeg

[palkod]

Draft Status

Draft - team will hold off on page creation

Category

Segmentation / Classification / Landmarking

Key Investigators

• Csaba Pinter (EBATINCA, Spain)
• Daniel Palkovics (Semmelweis University, Hungary)
• Andres Diaz-Pinto (NVIDIA, UK)

Project Description

Majority of currently available deep learning (DL) cone-beam computed tomography (CBCT) segmentation models were trained on data of healthy, completely dentated patients. These models might not produce accurate segmentations of datasets with dentoalveolar hard tissue defects. Our group has perviously developed a Deep Learning-based model for the automatic segmentation of dental cone-beam computed tomography (CBCT) scans which was trained on CBCT images with dentoalveolar pathological processes [1][2]. The current model uses a two-staged SegResNet-based architecture from MONAILabel. Despite of the relatively low sample training data it produced sufficient accuracy (93% compared to semi-automatic segmentation). However, the model's robustness has to be improved. Using the MONAI Auto3DSeg framework and an enlarged training database the project aims to develop an improved model for the automatic segmentation of dental CBCT scans present with dentoalveolar pathological processes.

Objective

1. Objective A. Develop an improved and robust automatic segmentation model for dental CBCT scans

Approach and Plan

1. Established and enlarged training database with uniformly annotated CBCT data.
2. Decide for an adequate network framework and architecture (MONAI Auto3DSeg?)
3. Come up with an initial configuration of the chosen architecture (stages, options, pre- and post-processing)
4. Perform preliminary training on the available data

Progress and Next Steps

We have previously trained a two-stage SegResNet-based model for the automatic segmentation of dental CBCT scans. The project was initiated at the 36th project week.

Illustrations

Two-stage SegResNet architecture

A: semi-automatic segmentation, B: deep learning segmentation

Background and References

1. Hegyi, A., Somodi, K., Pintér, C., Molnár, B., Windisch, P., García-Mato, D., Diaz-Pinto, A., & Palkovics, D. (2024). Mesterséges intelligencia alkalmazása fogászati cone-beam számítógépes tomográfiás felvételek automatikus szegmentációjára [Automatic segmentation of dental cone-beam computed tomography scans using a deep learning framework]. Orvosi hetilap, 165(32), 1242–1251. https://doi.org/10.1556/650.2024.33098
2. Palkovics, D., Hegyi, A., Molnar, B., Frater, M., Pinter, C., García-Mato, D., Diaz-Pinto, A., & Windisch, P. (2025). Assessment of hard tissue changes after horizontal guided bone regeneration with the aid of deep learning CBCT segmentation. Clinical oral investigations, 29(1), 59. https://doi.org/10.1007/s00784-024-06136-w

[sjh26]

@palkod Is this ready to submit?

[cpinter]

I'm pretty sure it is. Thank you!