Encrypted login | home

Program Information

Machine Learning and Texture Analysis for Assessing Spatial Tumor Response Based On Daily CTs During Radiation Therapy for Pancreatic Adenocarcinoma

no image available
D Schott

D Schott1*, W Hall1 , T Schmidt2 , X Chen1 , S Klawikowski1 , G Noid1 , P Knechtges1 , B Erickson1 , X Li1 , (1) Medical College of Wisconsin, Milwaukee, WI, (2) Marquette University, Milwaukee, WI


MO-RPM-GePD-JT-4 (Monday, July 31, 2017) 3:45 PM - 4:15 PM Room: Joint Imaging-Therapy ePoster Theater

Purpose: Develop a method using a machine learning classifier based on CT textures to identify spatial change of gross tumor volume (GTV) from daily CTs during the course of chemoradiation therapy (CRT) for pancreatic cancer.

Methods: Daily CT acquired using an in-room CT during CRT for pancreatic cancer patients along with pre- and post-CRT MRIs were analyzed. Maps of first order textures (mean, SD, entropy, skewness, and kurtosis) and fine second order textures (Grey Level Co-Occurrence matrix) were created from daily CTs. The pre- and post-CRT (pre-surgery) MRIs were registered to the first and last daily CTs and used to define the GTVs (pre-GTV and post-GTV), regions of fibrosis and/or necrosis (pre-ROFN and post-ROFN), and normal pancreatic tissues (pre-NPT and post-NPT). The classifier was trained to sort voxels in the pancreas into GTV, ROFN, and NPT from the first and last CT. The optimum combination of textures was defined by reiterating the training process with all possible combinations. The trained classifier was then used to analyze all daily CTs to identify the regions of interesting regions (ROI).

Results: For the 2 representative cases studied (well differentiated and poorly differentiated), the classifier was trained with an accuracy of 99% and 93% individually, and 98% combined. The Dice Coefficients between the ROIs identified by the trained classifier and actual ROIs on the last daily CT (e.g., post- GTV, ROFN, and NPT) were 71% well differentiated and 50% poorly differentiated. Large sets of patient data are required to improve the performance of the classifier.

Conclusion: During CRT for pancreatic cancer, the treatment-induced spatial tumor changes can be determined from daily CTs by using the machine learning classifier trained with CT textures from large cohorts of patients.

Contact Email: