Author: Laila A Gharzai, Bharat B Mittal, Poonam Yadav 👨🔬
Affiliation: Northwestern Feinberg School of Medicine, Northwestern Memorial Hospital, Northwestern University Feinberg School of Medicine, Northwestern University Feinberg School of Medicine 🌍
Purpose: Multiple studies have shown the increasing role of deep learning in segmenting regions of interest. This work presents the feasibility of auto-segmenting the critical structures for head and neck cancer radiation therapy using UNET-based architecture.
Methods: Diagnostic contrast-enhanced CT from 340 head/neck cancer patients were used for supervised learning and segmentation of the brain, brainstem, lenses, eyes, mandible, parotid glands, and spinal cord. A domain expert contoured the learning dataset. Deep learning architectures, such as VGG16-based UNET and generic UNET, were employed for training and validating the scans. Multiple transformations were applied on each iteration of the input data for data augmentation purposes. A total of 71.2M trainable parameters were generated for this work, and standard statistical evaluation metrics were used to evaluate the result after 300 epochs of supervised learning.
Results: Both the models showed higher accuracy (higher than 0.93) for the mandible, brain, and brainstem. VGG-based architecture achieved a precision score of 0.94, while 0.84 for the standard UNET. The dice coefficient for all the OARs for VGG16 outcomes was 0.90 to 0.95 (except for lenses, i.e., 0.87), while the range for UNET-based results was 0.88 to 0.93. The average Jaccard similarity coefficient of VGG16-based architecture was 6.5% higher than the standard UNET model. The aggregate structural metric results for VGG16 was 0.9274, while for generic UNET, it was 0.89. For all the OARs except the eye lenses, the augmented VGG16 UNET model resulted in a > 10% reduction in mean absolute error.
Conclusion: VGG-16 Unet outperformed UNET due to its ability to extract higher-order features. Standard evaluation metrics indicate the auto-segmentation of critical OARs in H & N cancer radiation therapy is feasible. Future work involves enhancing the network's performance and including other H&N OARs beyond the current six.