Fully automatic multi-organ segmentation for head and neck cancer radiotherapy using shape representation model...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Intensity modulated radiation therapy (IMRT) for head and neck (H&N) cancer requires accurate organ-at-risk (OARs) delineation.
• Manual segmentation is time-consuming and inconsistent.
• Existing automated methods struggle with anatomical variations and low contrast.

Purpose of the Study:

• Develop a novel automated H&N OARs segmentation method.
• Combine a fully convolutional neural network (FCNN) with a shape representation model (SRM).
• Improve accuracy and efficiency in OARs segmentation for H&N cancer treatment.

Main Methods:

• Trained SRM and FCNN on manually segmented H&N CT scans.
• SRM learned latent shape representations, constraining FCNN training.
• Delineated nine OARs (brainstem, optic chiasm, mandible, etc.) on unseen CT images.
• Evaluated using Dice Similarity Coefficient (DSC), PPV, SEN, ASD, and 95%SD.

Main Results:

• Achieved high average DSC for mandible (0.937) and brainstem (0.870).
• Demonstrated superior performance compared to atlas, statistical shape, and patch-wise CNN methods.
• Automated segmentation of 9 OARs took an average of 9.5 seconds per scan.

Conclusions:

• The proposed deep neural network segmentation method is effective for multi-organ segmentation on H&N CT scans.
• Incorporating shape priors (SRM) enhanced segmentation accuracy and robustness.
• The method offers competitive performance and reduced segmentation time compared to state-of-the-art techniques.