Anatomically consistent CNN-based segmentation of organs-at-risk in cranial radiotherapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate segmentation of organs at risk (OAR) is essential for radiotherapy planning to minimize side effects.
• Current segmentation methods can be computationally intensive and may lack ground truth data for all structures.

Purpose of the Study:

• To develop and validate a deep learning-based method for segmenting multiple OARs in the head from MRIs.
• To address computational costs and missing ground truth data in OAR segmentation.

Main Methods:

• A deep learning approach for end-to-end segmentation of eight head OARs (eye, lens, optic nerve, optic chiasm, pituitary gland, hippocampus, brainstem, brain).
• An efficient algorithm for training neural networks on multiple, non-exclusive classes.
• A postprocessing graph-based algorithm to enforce anatomical consistency, particularly for optic nerve connectivity.

Main Results:

• Quantitative cross-validation on 44 MRIs showed mean segmentation distances ranging from 0.1 to 0.7 mm.
• Qualitative assessment by a radiotherapist on 50 independent MRIs found 96% of segmentations acceptable for radiotherapy planning.
• The method demonstrated high accuracy and anatomical consistency for OAR segmentation.

Conclusions:

• The proposed deep learning method provides accurate and efficient segmentation of head OARs for radiotherapy planning.
• The system's ability to handle multiple OARs and enforce anatomical consistency enhances its clinical utility.
• This AI-driven approach shows significant promise for improving the safety and precision of radiotherapy treatments.