ESTRO 2024 - Abstract Book

S4957

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Fig.1 A diagram illustrating the main workflow.

Results:

Table 1 presents the test set AUC value of the optimal multivariate logistic regression model per endpoint, chosen by their mean AUC values from the training set. For grade ≥ 1 haemorrhage, the top model combining all the available features, yield a test AUC of 0.700 (95% CI 0.682 - 0.718). Notably, the planned dose parameters were discarded by the elastic net during training. The model incorporating planned dose parameters and daily MVCT radiomics excelled in predicting grade ≥ 2 haemorrhage, achieving an AUC of 0.820 (95% CI 0.806 - 0.834). Grade ≥ 1 proctitis had its highest test AUC at 0.565 (95% CI 0.538 - 0.593) using all available features, although accumulated dose parameters were discarded after univariate analysis. The model using only planning CT radiomic data performed well for grade ≥ 2 proctitis with an AUC of 0.776 (95% CI 0.754 - 0.798). Grade ≥ 1 stool frequency had the best model when planned dose parameters and planning CT radiomics were combined, obtaining an AUC of 0.627 (95% CI 0.613 - 0.642). Fig.2 shows the ROC curves for grade ≥ 1 and grade ≥ 2 haemorrhage using different feature combinations. For grade ≥ 2 haemorrhage, AUC values were consistently high with limited benefit from additional features compared to planned dose alone. However, grade ≥ 1 predictions improved with accumulated delivered dose and radiomic features, highlighting the role of radiomics which may be detecting indicators of underlying biological mechanism in toxicity prediction.

Highest p-value (vs. other feature combinations)

Endpoint

Model/Features

Test AUC

Grade ≥ 1 haemorrhage

AccD + CT + MVCT

0.700 (0.682 - 0.718)

< 0.001

Grade ≥ 2 haemorrhage

PlanD + MVCT

0.820 (0.806 - 0.834)

0.345

Grade ≥ 1 proctitis

PlanD + CT + MVCT

0.565 (0.538 - 0.593)

< 0.001

Grade ≥ 2 proctitis

CT

0.776 (0.754 - 0.798)

< 0.001