Program Information
Sensitivity of Radiomic Features to Image Noise and Respiratory Motion
K Lafata*, J Cai , C Wang , J Hong , C Kelsey , F Yin , Duke University Medical Center, Durham, NC
Presentations
SU-F-605-8 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: 605
Purpose: To investigate the effects of respiratory motion and image noise regarding radiomics analysis.
Methods: The sensitivity of 43 radiomic features to noise and breathing amplitude was investigated based on simulation and patient data. Extracted features from both phantom and patient CT data included: (1) Morphological-based, (2) Intensity-based, (3) Fine-Texture-based, and (4) Coarse-Texture-based. First, 16 dynamic respiratory environments were simulated with breathing amplitudes ranging from 0-to-30mm. Noise was sequentially added to the environment, producing 16 additional images with SNR ranging from 16-to-3. Features were extracted from each simulated scenario, and characterized by their change from baseline conditions. Next, 31 patients were retrospectively identified as meeting the following data availability requirements: existing free-breathing-CT and 4DCT images, and known tumor histology. Three feature-spaces were produced based on free-breathing, Average-Intensity-Projection (AIP), and End-of-Exhalation (EOE) phase images. Concordance correlation coefficients(CCC) were used to quantify feature-space variability between 3DCT and 4DCT acquisition. Lastly, a task-based approach was used to demonstrate how feature-space variability can ultimately influence radiomics end-points: Logistic regression algorithms were developed and independently trained to classify tumor histology based on the different feature-spaces. Area-under-the-curve (AUC) was used to evaluate classification results.
Results: Simulation results demonstrated strong linear dependences (p>0.95) between respiratory motion and morphological features (5-of-6), as well as between SNR and texture features (9-of-36). Regarding patient data, 40% of features demonstrated high CCC agreement (CCC>0.8) when comparing FB-to-EOE, compared to only 32% for FB-to-AIP. Histology model performance was directly proportional to SNR, and inversely proportional to temporal resolution. Statistically significant AUC values were 0.67 and 0.63 and 0.52 for AIP, free-breathing, and EOE feature-spaces, respectively. SNR was statistically significant between acquisition type, and appeared to be the primary factor in driving model performance.
Conclusion: Radiomics features are sensitive to respiratory motion and image noise, which can significantly influence task-based machine learning end-points.
Funding Support, Disclosures, and Conflict of Interest: Varian Medical Systems
Contact Email: