Computed tomography-based prediction of early recurrence risks with estimating individual times to recurrence for lung cancer patients prior to radiotherapy

Abstract

Background and Purpose

Current radiomics approaches have limitations in predicting recurrence with the time gap between radiomic features prior to treatment and individual recurrence risks following treatment. This study aimed to develop computed tomography (CT)-based prediction models of early recurrence risks by estimating individual times to recurrence (TTRs) in patients with non-small cell lung cancer (NSCLC) prior to stereotactic ablative radiotherapy (SABR).

Materials and methods

A total of 143 patients with stage I-II NSCLC treated with SABR were enrolled. Planning CT images were used to construct internal (n = 125; 43 recurrent, 56 censored, and 26 recurrence-free within 5 years) and external (n = 18; 15 recurrent and 3 recurrence-free) datasets. Reference TTR data from 124 recurrent patients were used to estimate virtual TTRs of 56 censored patients using a Bayesian approach with the Markov chain Monte Carlo algorithm. Three types of regression models were developed to estimate the TTR based on significant CT image features associated with early recurrences. The predictions of 1-year and 2-year recurrences based on the estimated TTRs were evaluated using the areas under the receiver operating characteristic curves (AUCs).

Results

The best prediction models achieved AUCs of 0.926 and 0.763 for 1-year and 0.869 and 0.883 for 2-year recurrence in the internal and external tests, respectively.

Conclusion

The proposed CT-based model demonstrated improved performance for predicting 1- and 2-year recurrences by directly predicting TTRs. This study could bridge the time gap between pretreatment radiomic features and individual recurrence risks.