A slice classification model-facilitated 3D encoder-decoder network for segmenting organs at risk in head and neck...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Deep learning models for segmenting organs at risk (OARs) in head and neck (H&N) cancer computed tomography (CT) images face challenges with class imbalance, leading to inaccuracies.
• This imbalance causes false positives on irrelevant slices and increases computation time, impacting clinical workflow.

Purpose of the Study:

• To develop and validate a novel two-step deep learning network to address the class-imbalance problem in H&N cancer OAR segmentation.
• To improve the accuracy and efficiency of OAR delineation in CT images.

Main Methods:

• A slice classification model was developed to categorize CT slices into six craniocaudal directions.
• Target OAR categories were then directed to specialized 3D encoder-decoder segmentation networks.
• The model was trained and validated on datasets comprising 120 (training), 30 (validation), and 20 (testing) patients.

Main Results:

• The slice classification model achieved an average accuracy of 95.99%.
• High Dice similarity coefficients and low 95% Hausdorff distances were reported for various OARs, including the eyes, brainstem, and mandible.
• The total segmentation time was significantly reduced to 40.13 seconds.

Conclusions:

• The developed slice classification model-facilitated 3D encoder-decoder network demonstrates superior accuracy and efficiency for H&N cancer OAR segmentation.
• This approach effectively mitigates class-imbalance issues, promising to reduce radiation oncologists' workload.