Sensitivity of Radiomic Features to Acquisition Noise and Respiratory Motion

Abstract

The sensitivity of radiomics features to different imaging parameters is poorly understood. The purpose of this work was to investigate the effects of motion and noise regarding radiomics analysis of dynamic cancers. The sensitivity of 43 radiomic features to image noise and breathing amplitude was investigated based on simulation and patient data. Extracted features were classified into 1 of 4 categories: (1) Morphological, (2) Intensity, (3) Fine Texture, and (4) Coarse Texture. First, a series of simulations were used to study feature sensitivity within a controlled tumor environment. Using the Extended Cardiac-Torso code, 16 dynamic respiratory environments were simulated with breathing amplitudes ranging from 0 to 30 mm. Noise was sequentially added by convolving the tumor with various Gaussian kernels, producing 16 simulated images with SNR values 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 3 data availability requirements: an existing free breathing (FB) CT, an existing 4DCT, and a known tumor histology. For each patient, 3 feature spaces were produced based on each FB image, Average Intensity Projection (AIP) image, and End of Exhalation (EOE) phase image. Paired sample t-tests and concordance correlation coefficients (CCC) were used to determine which features demonstrated significant variability between 3D and 4D image acquisition. Finally, a task-based approach was used to demonstrate how such feature 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. Simulation results demonstrated strong linear dependences (P > 0.95) between respiratory motion and morphological features (5/6), as well as between SNR and texture features (9/36). Regarding patient data, 40% of features demonstrated high CCC agreement (CCC >0.80) when comparing FB to EOE, compared to only 32% between FB and AIP. Similarly, 32% of features demonstrated significant P-values when comparing FB to AIP, compared to only 21% between FB and EOE. 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, FB, and EOE feature spaces, respectively. SNR values were statistically significant between acquisition type, and appeared to be the primary factor in driving model performance. Radiomic feature sensitivity to motion and noise was investigated using both simulation and patient data. By training acquisition specific machine learning models, we demonstrated that such variability can greatly influence task-based radiomics end-points, such as histological classification.