時間的画像を用いた新しい累積モデリング手法：概念提案と複数疾患に対する臨床的検証

To propose and cross-validate a novel accumulative predicting method using temporal data of image-guided-radiotherapy.

CBCTacquired before the first treatment, and thei^thweek thereafter were denoted as CBCT₀and CBCT_i(i = 1,2,…). Temporal changes of 560 radiomics features were calculated (Delta-RF_i,i = 1,2,…) and stacked to a novel accumulative delta-radiomics feature (Delta-RF_accu).The predicting performance of the Delta-RF_i(i = 1,2,…,accu) signatures was compared using univariate Logistic regression. The Delta-RF_isignatures and important clinical/dosimetric predictors were incorporated into multivariate Logistic regression models. Two clinical tasks were used to test the generalizability: the prediction of radiation pneumonitis (RP) risk, and the pathologic complete response (pCR) for rectal patients.

The AUC of the Delta-RF_accusignature was 0.82 ± 0.09 for RP prediction, significantly higher than the conventional Delta-RF₁, Delta-RF₂, and Delta-RF₃(0.70 ± 0.09, 0.66 ± 0.15 and 0.70 ± 0.13, respectively). The improvement was reproducible in pCR prediction. After incorporating Delta-RF_accuwith important clinical and dosimetric factors, the AUC was further increased to 0.85 ± 0.10 and 0.87 ± 0.06, respectively in RP and pCR prediction tasks. Using temporal CBCT images of relatively lower quality, the AUC of the proposed method was comparable if not superior than the previous studies using CT or MRIin the prediction tasks.

A novel accumulative modelling method using temporal images was proposed and the superior predicting ability was validated on multiple diseases. The proposed method and the clinical models can be used to support patient-specific decision-making, reducing radiation injury or avoiding unnecessary surgery.