A convolutional neural network algorithm for automatic segmentation of head and neck organs at risk using deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Accurate segmentation of head and neck organs at risk (OARs) is crucial for radiation therapy planning.
• Current single-task learning methods may not fully leverage shared anatomical information across OARs.
• Deep learning models like U-Net have shown promise but can be computationally intensive and require extensive task-specific training.

Purpose of the Study:

• To introduce and evaluate a lifelong learning-based convolutional neural network (LL-CNN) algorithm for automatic OAR segmentation.
• To compare the performance of LL-CNN against established methods such as 2D-UNet, 3D-UNet, single-task CNN (ST-CNN), and multitask CNN (MT-CNN).
• To demonstrate LL-CNN's superiority in prediction accuracy for head and neck OAR segmentation.

Main Methods:

• A multitask learning framework was employed to train a shared network on twelve head and neck OARs simultaneously.
• The trained network's final layer was adapted for single-task transfer learning, with subsequent training on individual OARs using early stoppage.
• Performance was quantified using Dice score and root-mean-square error (RMSE), with comparisons made against manually delineated contours (gold standard).
• Training and validation adhered to Kaggle competition standards, utilizing 160 patients for training, 20 for internal validation, and 20 for final testing.

Main Results:

• LL-CNN achieved higher average Dice coefficients and lower RMSE compared to 2D-UNet, 3D-UNet, ST-CNN, and MT-CNN.
• The LL-CNN model required approximately 72 hours for training on two Nvidia 1080Ti GPUs using distributed learning.
• Prediction time for all 12 OARs using LL-CNN was around 20 seconds, comparable to the fastest alternative methods.

Conclusions:

• The LL-CNN algorithm demonstrates superior prediction accuracy for segmenting head and neck organs at risk compared to all evaluated alternative algorithms.
• Lifelong learning offers a robust and accurate approach for complex medical image segmentation tasks.
• LL-CNN presents a promising advancement in automated OAR segmentation, potentially improving radiotherapy planning efficiency and precision.