Room: Room 202
Purpose: To expedite the contouring process for MR image guided adaptive radiotherapy (MR-IGART), a deep densely connected convolutional neural network (DeepDenseNet) is proposed to automatically segment liver, kidneys, stomach and large bowel in ViewRay 3D MR images.
Methods: Datasets of 120 patients were collected retrospectively. Treatment sites included pancreas, liver, stomach, adrenal gland and prostate. DeepDenseNet was trained using 100 datasets and tested on the remaining 20 datasets. DICE coefficient, median Hausdorff distance (HD) and max HD were calculated to evaluate the segmentation accuracy. To demonstrate the clinical relevance, the time cost to manually contour the organs with and without the help of the DeepDenseNet was compared.
Results: DeepDenseNet could segment the organs with good accuracy and speed. For the 20 testing patients, the average DICE coefficients were 0.87, 0.88, 0.95 and 0.92 for large bowel, stomach, liver and kidneys respectively. The average median HDs were 2.59 mm, 2.72 mm, 2.34 mm and 2.31 mm respectively. The manual contour timing comparison showed that manual contouring following the DeepDenseNet automatic segmentation results was 5, 3 and 2 times faster than manual contouring from scratch for all three studied cases.
Conclusion: DeepDenseNet can segment multiple abdominal organs automatically in ViewRay MRIs with good accuracy. It is useful to expedite the manual contouring process for MR-IGART.