Recurrent Convolutional Neural Networks for 3D Mandible Segmentation in Computed Tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Classic encoder-decoder-based convolutional neural networks (EDCNNs) struggle with accurate mandible segmentation in CT scans.
• Noise and metal artifacts in CT scans further challenge the precise segmentation of mandibular structures like condyles and coronoids.
• Existing EDCNN approaches often overlook the crucial anatomical connectivity of the mandible.

Purpose of the Study:

• To propose a novel 3D convolutional neural network (CNN) based approach for accurate mandible segmentation.
• To address the limitations of EDCNNs in segmenting detailed mandibular anatomical structures affected by noise and artifacts.
• To develop a method that accounts for the anatomical connectivity of the mandible during segmentation.

Main Methods:

• A novel Recurrent CNN Segmentation (RCNNSeg) approach was developed, utilizing recurrent neural networks to maintain anatomical connectivity.
• RCNNSeg processes complete 3D CT scans, unlike traditional methods requiring 2D slices or 3D patches.
• The method was evaluated on 109 local head and neck CT scans and 40 PDDCA public dataset scans, using Dice Similarity Coefficient (DSC), Average Symmetric Surface Distance (ASD), and 95% Hausdorff Distance (95HD).

Main Results:

• RCNNSeg demonstrated superior performance over EDCNN-based approaches on both local and PDDCA datasets.
• The method achieved high accuracy with an average DSC of 97.48% and ASD of 0.2170 mm on the local dataset.
• On the PDDCA dataset, RCNNSeg achieved an average DSC of 95.10% and ASD of 0.1367 mm, outperforming state-of-the-art methods.

Conclusions:

• The proposed RCNNSeg method provides more accurate automated mandible segmentations compared to classic EDCNN techniques.
• RCNNSeg effectively learns and utilizes spatially structured information for improved mandible segmentation.
• The approach shows significant potential for automatic and precise mandible segmentation in clinical applications.