Impact of experimental design on PET radiomics in predicting somatic mutation status

Abstract

Purpose

PET-based radiomic features have demonstrated great promises in predicting genetic data. However, various experimental parameters can influence the feature extraction pipeline, and hence, Here, we investigated how experimental settings affect the performance of radiomic features in predicting somatic mutation status in non-small cell lung cancer (NSCLC) patients.

Methods

348 NSCLC patients with somatic mutation testing and diagnostic PET images were included in our analysis. Radiomic feature extractions were analyzed for varying voxel sizes, filters and bin widths. 66 radiomic features were evaluated. The performance of features in predicting mutations status was assessed using the area under the receiver-operating-characteristic curve (AUC). The influence of experimental parameters on feature predictability was quantified as the relative difference between the minimum and maximum AUC (δ).

Results

The large majority of features (n = 56, 85%) were significantly predictive for EGFR mutation status (AUC ≥ 0.61). 29 radiomic features significantly predicted EGFR mutations and were robust to experimental settings with δ_Overall < 5%. The overall influence (δ_Overall) of the voxel size, filter and bin width for all features ranged from 5% to 15%, respectively. For all features, none of the experimental designs was predictive of KRAS+ from KRAS− (AUC ≤ 0.56).

Conclusion

The predictability of 29 radiomic features was robust to the choice of experimental settings; however, these settings need to be carefully chosen for all other features. The combined effect of the investigated processing methods could be substantial and must be considered. Optimized settings that will maximize the predictive performance of individual radiomic features should be investigated in the future.

Introduction

Functional positron emission tomography imaging with the glucose analog ¹⁸F-fluorodoxyglucose (¹⁸F-FDG PET) is widely used for diagnosis and staging in oncology. ¹⁸F FDG PET plays an increasingly important role in the assessment of treatment response [1] and planning of radiotherapy [2]. Furthermore, ¹⁸F FDG PET imaging can capture the different metabolic phenotypes that exist between tumors [3], [4]. Quantification of these phenotypes may improve tumor characterization for individualized therapy [4], [5]. Radiomics extracts an atlas of features that describe the tumor phenotype from standard of care images through the utilization of advanced mathematical algorithms that quantify the relationship between image voxels [6], [7]. Recently, PET radiomic features have demonstrated tremendous promise in predicting clinical outcomes [8], [9], treatment response [10], [11], and genomic data [3] in various malignancies.

As part of the feature extraction pipeline, PET images undergo two image processing methods [5]. First, the voxel sizes are resampled. Thus far, radiomic studies have relied on retrospective PET imaging data, which are often acquired on different scanners and are reconstructed differently, leading to variability in the voxel size [12], [13]. Therefore, for radiomic analysis, many research groups resample the voxel size with an interpolation filter in order to obtain a uniform size across all patients [5], [14]. Second, the image intensity of PET images are often discretized (or binned) into a limited range of intensity values to reduce image noise [5] and increase feature computational efficiency [6].

The values and methods chosen for these image processing steps are often selected without clear justification or are simply not mentioned in many radiomic studies. Therefore, it is unclear whether these parameters impact the performance of PET radiomic features in predicting outcomes or status and what parameters will maximize the predictability of the features. Researchers have reported that the quantitative value of PET radiomic features is sensitive to the choice of discretization methods [15], [16]. However, the effects of the resampled voxel size and interpolation filter on PET feature quantification have not been investigated. Furthermore, the impact of these settings on the ability of radiomic features to predict clinically valuable information has not been reported. With the overall goal of the clinical implementation of radiomics, it is important to study the impact of these parameters on the predictive performance of radiomic features.

Due to the high incidence and recurrence rate of non-small cell lung cancer (NSCLC) [17], there has been considerable interest in applying radiomics to characterize the disease in order to improve prognosis and better tailor treatments for individual patients [4]. Initiation and progression of NSCLC tumors are driven by somatic mutations in key oncogenes (e.g. epidermal growth factor receptor (EGFR) and Kirsten rat sarcoma viral oncogene homolog (KRAS)) [18]. Early prediction of the oncogene mutation status may provide information for individualized therapy for NSCLC patients [18]. PET-based radiomic features have shown to be significantly associated with NSCLC mutation status [3]. Furthermore, unlike prediction of treatment outcomes (e.g. overall survival), the prediction of somatic mutations is not confounded by treatment methods as they are an intrinsic characteristic of the tumor. In this study, we investigated how various experimental settings influence the feature quantification and performance of radiomic features in predicting somatic mutations.