(Monday, 7/30/2018) 1:15 PM - 1:45 PM
Room: Exhibit Hall | Forum 6
Purpose: Develop a method to identify the spatial change of gross tumor volume (GTV) using a machine learning classifier based on CT textures from daily CTs during the course of radiation therapy (RT).
Methods: The proposed method is demonstrated on daily CTs acquired using an in-room CT during RT for three patients (lung, breast, and pancreas), representing high, medium, and low image contrast between tumor and surrounding tissues, respectively. Maps of first-order and second-order image textures were extracted from GTVs on daily CTs. The pre- and post-RT images (RT planning MRI and/or contrast CT and follow up MRI or CT with visible GTV) were registered to the first and last daily CTs and used to define the GTVs (pre-GTV and post-GTV) and normal tissues. With GTV ground truth data at the first fractions, the classifier was trained using 50% of voxels to select appropriate classifier and textures. The classifier was then refined with the remaining voxels by optimizing combinations of the selected textures using a reiterative process. On a daily CT, the trained classifier was used to identify the GTV of the day. The model predicted GTV was compared with the ground truth GTV at the last daily CT.
Results: For the three types of tumor sites studied, the classifier was trained with 99% accuracy for lung and breast, and 95% for pancreas. The Dice Coefficients between the GTVs identified by the trained classifier in the last daily CT and post-GTVs ranged from 84% higher contrast (lung) to 70% low contrast (pancreas).
Conclusion: A method to assess spatial tumor changes during RT is developed based on image texture analysis of daily CTs acquired during RT using a machine learning classifier. The method, once validated with a large cohort of patients, may be used to delineate GTV for adaptive RT.
Keywords
Tissue Characterization, Quantitative Imaging, CT
Taxonomy
Contact Email
