Optimizing the CBCT Segmentation Pipeline with Intuition-Guided Processing

Qingyu Kuang

Keywords: CBCT Segmentation, Deep Learning, medical images

Abstract: In the past, general medical image models attracted considerable research interest. However, since medical imaging modalities vary widely and often fundamentally differ from RGB images, applying a general segmentation framework to specific tasks usually requires further optimization to achieve satisfactory performance. Cone-beam computed tomography (CBCT) is a commonly used medical imaging technique in dentistry. Optimizing the segmentation process for CBCT images can greatly enhance the effectiveness of computer-aided diagnostic systems in dental applications. In this work, we analyzed the ToothFairy3 dataset and proposed improvements to the nnU-Net framework. While preserving the auto-configuration capabilities of nnU-Net, we introduced targeted optimizations across the data preprocessing pipeline, network architecture, inference process, and postprocessing strategies to enhance performance for the CBCT multi-class segmentation task. Furthermore, the trained multi-class segmentation model can be integrated with user click prompts to train an interactive segmentation model. These modifications collectively reduced inference time, improved model effectiveness, and increased practical applicability.

Changes Summary:

Thank you very much to the reviewers for your valuable comments.
All requested additions have been added, including the training parameter settings and the loss functions.
We have also provided further details on the model training process, including the motivation for “dimension expansion”. However, this setting involves a certain degree of subjectivity, as it was determined based on the visual appearance of the images. And different parameter ranges during edge detection can also lead to different results; therefore, we only report the final results on the test set. In future work, we plan to validate these settings using alternative methods.

Main Tex File: samplepaper.tex

Confirm Latex Only: true

Submission Number: 18