A machine learning based delta-radiomics process for early prediction of treatment response of pancreatic cancer

Abstract

Changes of radiomic features over time in longitudinal images, delta radiomics, can potentially be used as a biomarker to predict treatment response. This study aims to develop a delta-radiomic process based on machine learning by (1) acquiring and registering longitudinal images, (2) segmenting and populating regions of interest (ROIs), (3) extracting radiomic features and calculating their changes (delta-radiomic features, DRFs), (4) reducing feature space and determining candidate DRFs showing treatment-induced changes, and (5) creating outcome prediction models using machine learning. This process was demonstrated by retrospectively analyzing daily non-contrast CTs acquired during routine CT-guided-chemoradiation therapy for 90 pancreatic cancer patients. A total of 2520 CT sets (28-daily-fractions-per-patient) along with their pathological response were analyzed. Over 1300 radiomic features were extracted from the segmented ROIs. Highly correlated DRFs were ruled out using Spearman correlations. Correlation between the selected DRFs and pathological response was established using linear-regression-models. T test and linear-mixed-effects-models were used to determine which DRFs changed significantly compared with first fraction. A Bayesian-regularization-neural-network was used to build a response prediction model. The model was trained using 50 patients and leave-one-out-cross-validation. Performance was judged using the area-under-ROC-curve. External independent validation was done using data from the remaining 40 patients. The results show that 13 DRFs passed the tests and demonstrated significant changes following 2-4 weeks of treatment. The best performing combination differentiating good versus bad responders (CV-AUC = 0.94) was obtained using normalized-entropy-to-standard-deviation-difference-(NESTD), kurtosis, and coarseness. With further studies using larger data sets, delta radiomics may develop into a biomarker for early prediction of treatment response.

Keywords: Cancer imaging; Cancer models; Mathematics and computing; Prognostic markers; Tumour biomarkers.

Fig. 1

DRFs’ Spearman correlations. A sample of the correlation histogram with p-values is shown on the left, Spearman ranking (middle) for some of the DRFs, and an example of a Spearman Correlation heatmap for some of the DRFs (right)

Fig. 2

DRFs as a function of motion artifact. A comparison of two DRFs, cluster prominence (left), and texture strength (right), along with the t test p-value in different weeks during CRT between the low and high motion groups, and the difference in coefficient of variance (COV) between high and low motion for selected DRFs (right) with 5–10% levels indicated with cluster prominence and texture strength highlighted

Fig. 3

DRFs t test and distributions. Boxplots and corresponding t test p-value for (a) average DRFs of all fractions for maximum to mean ratio (MaxMean) feature indicating no significant difference between the two response groups and for the contrast feature showing significant differences. The presented boxplots show the median and interquartile range for each response group, and the diamond data point in the middle represents the mean of the group, (b) distribution of good- and bad-response group including all patients and all fractions within each response group for Coarseness, Kurtosis, NESTD, and contrast

Fig. 4

Fractional changes of the contrast feature. Comparisons of the boxplots and p-values of daily DRFs of the contrast feature for all patients in a response group between the two response groups, showing (a) all daily DRFs for both groups plotted together and (b) daily DRFs for the two response groups plotted for each fraction

Fig. 5

3D scatter plots for the best performing feature combination. The weekly DRFs for weeks 2–4 for 50 good responders (150 data points) and the 40 bad responders (120 data points) for the best performing features combinations (kurtosis–NESTD–coarseness) is displayed

Fig. 6

A general delta-radiomics process. Including extracting delta-radiomics features and building machine-learning models for treatment outcome prediction

Fig. 7

An example of NESTD map generation. a entropy map, b STD map, c the resultant NESTD map, and d the NESTD map overlaid on the original CT image