Do Dosiomic Features Extracted From Planned 3-Dimensional Dose Distribution Improve Biochemical Failure-Free Survival Prediction: an Analysis Based on a Large Multi-Institutional Data Set

Abstract

Purpose: The objective of this work was to investigate whether including additional dosiomic features can improve biochemical failure-free survival prediction compared with models with clinical features only or with clinical features as well as equivalent uniform dose and tumor control probability.

Methods and materials: This retrospective study included 1852 patients who received diagnoses of localized prostate cancer between 2010 and 2016 and were treated with curative external beam radiation therapy in Albert, Canada. A total of 1562 patients from 2 centers were used for developing 3 random survival forest models: Model A included only 5 clinical features; Model B included 5 clinical features, equivalent uniform dose, and tumor control probability; and Model C considered 5 clinical features and 2074 dosiomic features derived from the planned dose distribution of the clinical target volume and planning target volume with further feature selection to determine prognostic features. No feature selection was performed for models A and B. Two hundred ninety patients from another 2 centers were used for independent validation. Individual model-based risk stratification was examined, and the log-rank tests were performed to test statistically significant differences between the risk groups. The 3 models' performances were evaluated using Harrell's concordance index (C-index) and compared using one-way repeated-measures analysis of variance with post hoc paired t test.

Results: Model C selected 6 dosiomic features and 4 clinical features to be prognostic. There were statistically significant differences between the 4 risk groups for both training and validation data sets. The C-index for the out-of-bag samples of the training data set was 0.650, 0.648, and 0.669 for models A, B, and C, respectively. The C-index for the validation data set for models A, B, and C was 0.653, 0.648, and 0.662, respectively. Although gains were modest, Model C was statistically significantly better than models A and B.

Conclusions: Dosiomics contain information beyond common dose-volume histogram metrics from planned dose distributions. Incorporation of prognostic dosiomic features in biochemical failure-free survival outcome models can lead to statistically significant although modest improvement in performance.

Figure 1

Workflow in this study. The full patient data set was divided into a training data set (patients treated at 2 tertiary cancer centers) and a validation data set (patients treated at 2 regional cancer centers). For Model A, which included only clinical features, and Model B, which included clinical features as well as equivalent uniform dose (EUD) and tumor control probability (TCP), the training data set was used to develop the random survival forest models. For Model C, which examined clinical features and dosiomic features, additional feature selection steps were used before hyperparameter tuning. All 3 models’ discrimination and risk groups were evaluated on the training and the validation data set. Abbreviations: CTV = clinical target volume; EBRT = external beam radiation therapy; KM = Kaplan-Meier; PTV = planning target volume; ROI = region of interest; RSF = random survival forest.

Figure 2

Variable importance in Model C. The variable importance was calculated as the amount the C-index decreased when the variable values for all patients were randomly shuffled. The higher the importance, the most prognostic the variable was in predicting biochemical failure-free survival. Abbreviations: CTV = clinical target volume; GLSZM = gray level size zone matrix; PSA = prostate-specific antigen; PTV = planning target volume.

Figure 3

Kaplan-Meier survival curves for patients with planning target volume (PTV) wavelet HLH gray level size zone matrix (GLSZM) entropy feature less than or equal to the median and greater than the median of the training data set. The shaded region represents the 95% confidence interval. (a) For the training data set, a statistically significant difference (P < .01) was found between the 2 groups. (b) For the validation data set, there were no statistically significant differences (P = .4). Abbreviation: EBRT = external beam radiation therapy.

Figure 4

Kaplan-Meier survival curves for the low-, favorable intermediate-, unfavorable intermediate-, and high-risk groups. The shaded regions represent the 95% confidence internal. (a, b) Model A–based risk stratification for the training and the validation data set. (c, d) Model B–based risk stratification for the training and the validation data set. (e, f) Model C–based risk stratification for the training and the validation data set. Statistically significant differences (all P < .01) were identified among the 4 risk groups for all 6 cases.