Room: AAPM ePoster Library
Purpose: of the existing radiomics research of the head and neck (H&N) cancer extracts quantitative features to correlate with clinical endpoints. However, the prediction performance of local recurrence (LR) is still challenging. We aimed to develop and validate a 2.5 dimensional (2.5D) transfer learning based model for improving LR prediction in laryngeal cancer.
Methods: total of 48 laryngeal cancer patients with pre-treatment Computed Tomography (CT) from The Cancer Imaging Archive (TCIA) were included in this study (LR=37, No LR=11). 403 2D patches derived from the gross tumor volume (GTV) on the planning CT were extracted and refined by removing the overlapped bone regions and slices without tumor parenchyma as the dataset for model training. A convolutional neural network (CNN) model transferred from VGG19 was pre-trained on ImageNet, then fine-tuned on the tumor patches. Image augmentation and weighted loss function were employed to address the data imbalance. The group of LR prediction results based on a stack of 2D slices were combined through a majority vote to provide the final prediction for a given patient. The performance was evaluated using Area Under Curve (AUC), sensitivity and specificity in 8-folds cross validation in the cohort.
Results: conventional H&N radiomics model trained with tumors in both oropharynx and larynx yielded AUCs ranging from 0.48 to 0.65 in predicting LR of H&N cancer. In comparison, the proposed model achieved an average AUC=0.91, sensitivity=1, and specificity=0.81 for prediction of LR using an independent cohort of laryngeal cancer.
Conclusion: prediction power of radiomics features for LR is significantly enhanced for the subtype, laryngeal cancer, compared with LR prediction of H&N cancer including tumors in both oropharynx and larynx. Refining the GTV contour to get the homogeneous tumor region is essential for improving the performance of image-based prediction with limited dataset.
Image Analysis, CT, Pattern Recognition
TH- Response Assessment: Radiomics/texture/feature-based response assessment