MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Machine Learning Process for Delta Radiomics

H Nasief*, X Li , Froedtert Hospital and the Medical College of Wisconsin, Milwaukee, WI

Presentations

(Tuesday, 7/31/2018) 1:15 PM - 1:45 PM

Room: Exhibit Hall | Forum 3

Purpose: Change of radiomic features in longitudinal images, delta-radiomics, can be potentially used as an imaging bio-marker for treatment response. This study outlines a methodology for delta-radiomic analysis based machine learning algorithm.

Methods: The proposed methodology is demonstrated using daily CTs acquired during routine CT-guided chemo-radiation therapy (CRT) for 31 patients with pancreatic head cancer. Patients were divided into good- and poor-response groups based on their pathological responses, and then further divided into training (n=20) and testing (n=11) sets. Changes of 73 radiomic features between fractions were extracted from the segmented tumor on daily CTs. A new feature, called NESTD (defined as normalized entropy to std. difference) is introduced. A two-step machine learning algorithm with shallow and deep network was developed to generate predictive models. Training with the shallow network utilized a self-organization map to group features into classes based on similarity and to take into account that data are clustered in many dimensions. The deep network utilized bootstrapped samples generated based on the shallow network model and delta-radiomics, and a Bayesian regularization scheme to avoid overfitting. Due to limited data set, combination of two or 3 features were examined at a time. Cross-validation was performed using a leave-one-out method. Classification was based on the minimum Mahalnobios distance to the centroid of each class and performance was judged using the area under the ROC curve (AUC).

Results: Delta-radiomic features that have potential to distinguish the two response groups were identified. The best performing two feature combination (AUC=0.92) was obtained using Kurtosis and coarseness. The performance increased to 94% using Kurtosis, Coarseness and NESTD.

Conclusion: Machine learning process can identify appropriate delta-radiomic features predicting treatment response during CRT delivery for pancreatic cancer. With further studies based on large patient datasets, delta-radiomics may be developed as an imaging biomarker for adaptive RT.

Keywords

CT, Image-guided Therapy, Classifier Design

Taxonomy

TH- response assessment : Machine learning

Contact Email