Abstract 956P
Background
The objective of this study is to use a deep learning approach to develop a model that combines CT and dose images to predict the occurrence of radiation pneumonitis (RP) in lung cancer patients who received radical (chemo)radiotherapy.
Methods
CT images, radiation dose images and clinical parameters were obtained from 314 retrospectively-collected patients (training set) and 35 prospectively-collected patients (test-set-1) diagnosed with lung cancer between 2013 and 2019. Another 153 (60Gy group, test-set-2) and 89 (74Gy group, test-set-3) patients from the clinical trial RTOG 0617 were used for external validation. A ResNet architecture was applied as the backbone to develop a prediction model that incorporates CT and dose images. The CT and dose weights were adjusted to accommodate cohorts with different clinical settings or dose delivery patterns. The test sets were used to evaluate the model. Visual interpretation was implemented using a gradient-weighted class activation map (grad-CAM) to observe the area of model attention during the prediction process. To improve the ease of use of the model, ready-to-use online software was developed.
Results
The discriminative ability of the model was AUC 0.83, Accuracy 0.82, Sensitivity 0.70, Specificity 0.88 for test-set-1, AUC 0.55 for test-set-2, and AUC 0.63 for test-set-3. After adjusting the model with 40 patients from test-set-2 and test-set-3, the discriminatory power of test sets 2 and 3 improved to AUC 0.65 and AUC 0.70, respectively. Grad-CAM showed the regions of interest to the model that contribute to the prediction of RP. Table: 956P
Performance of proposed algorithm in the test sets
Dataset | AUC | Accuracy | Sensitivity | Specificity |
Test-set-1 | 0.834 (0.823- 0.906) | 0.824 (0.746- 0.812) | 0.703 (0.670 -0.770) | 0.875 (0.832-0.894) |
Test-set-2 | 0.552 (0.466-0.688) | 0.695 (0.572-0.818) | 0.413 (0.391-0.523) | 0.691 (0.612-0.837) |
Test-set-3 | 0.627 (0.531-0.724) | 0.664 (0.600-0.728) | 0.595 (0.456-0.735) | 0.682 (0.612-0.753) |
Test-set-2* | 0.650 (0.535-0.766) | 0.763 (0.627-0.907) | 0.578 (0.483-0.832) | 0.698 (0.662-0.975) |
Test-set-3* | 0.700 (0.625-0.762) | 0.712 (0.634-0.831) | 0.615 (0.563-0.855) | 0.733 (0.671-0.950) |
∗the asterisk indicates that the coefficients of CT and RD for this model are adjusted for each set of 40 patients.
Conclusions
A novel deep learning approach combining CT and dose images can effectively and accurately predict the occurrence of RP.
Clinical trial identification
This manuscript was prepared using data from Datasets [RTOG-0617] from the NCTN/NCORP Data Archive of the National Cancer Institute’s (NCI’s) National Clinical Trials Network (NCTN). Data were originally collected from clinical trial NCT number [NCT00533949] [RTOG-0617]. All analyses and conclusions in this manuscript are the sole responsibility of the authors and do not necessarily reflect the opinions or views of the clinical trial investigators, the NCTN, the NCORP or the NCI.
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.