Exhibit Hall | Forum 9
Purpose: To automate contouring of the elective clinical target volumes (CTVs) and organs at risk (OARs) for head and neck cancer (HNC) patients, treated on MR-Linac, through algorithmic segmentation of magnetic resonance (MR) images via deep convolutional neural networks (CNNs).
Methods: The MOMENTUM Study provides 38 MR T2-weighted scans of HNC patients. A clinician manually contoured the elective CTVs (lymph nodes), spinal cord (SC), parotid glands, and brainstem, referred to as C-ROIs. 34 MR scans together with the C-ROIs were used to train a CNN with a 2D U-Net architecture (Ronneberger et al, 2015). Data augmentation was performed through applying rotation up to range ±3°, zoom up to 10%, and vertical and horizontal shifts up to 10%. The algorithm was tested on the remaining 4 scans. Results were evaluated geometrically using the Dice similarity coefficient (DSC) and dosimetrically by evaluating the effect on the C-ROIs of optimizing new plans using algorithmically derived volumes applying the clinical goals used at our hospital to judge the suitability of clinical plans.
Results: Our results show mean DSC (± SD) of 0.54 ± 0.06 for left nodes, 0.60 ± 0.03 for right nodes, 0.68 ± 0.12 for SC, 0.80 ± 0.14 for left parotid, 0.83 ± 0.02 for right parotid, 0.85 ± 0.03 for brainstem. Dosimetrically we observed that all clinical goals remain satisfied except those judging the minimum dose level to be delivered to 95% and 99% of the elective CTV. Full segmentation executes in the order of 8 seconds.
Conclusion: This study shows the feasibility of using CNNs to automate the contouring of the OARs required for treatment planning of HNC patients on the MR-Linac. Results demonstrate difficulty in contouring the elective CTVs, however, larger datasets are becoming available which shall be used in combination with different approaches to improve contouring accuracy.
Segmentation, Radiation Therapy, MR