Deep Learning and Machine Learning Techniques for Automated PET/CT Segmentation and Survival Prediction in Head and Neck Cancer

Abstract

Background: Accurate prognostic stratification as well as segmentation of Head-and-Neck Squamous-Cell-Carcinoma (HNSCC) patients can be an important clinical reference when designing therapeutic strategies. We set to enable automated segmentation of tumors and prediction of recurrence-free survival (RFS) using advanced deep learning techniques and Hybrid Machine Learning Systems (HMLSs). Method: In this work, 883 subjects were extracted from HECKTOR-Challenge: ~60% of the total subjects were considered for the training and validation procedure, and the remaining subjects for external testing were employed to validate our models. PET images were registered to CT. First, a weighted fusion technique was employed to combine PET and CT information. We also employed Cascade-Net to enable automated segmentation of HNSCC tumors. Moreover, we extracted deep learning features (DF) via a 3D auto-encoder algorithm from PET and the fused image. Subsequently, we employed an HMLS including a feature selection algorithm such as Mutual Information (MI) linked with a survival prediction algorithm such as Random Survival Forest (RSF) optimized by 5-fold cross-validation and grid search. The dataset with DFs was normalized by the z-score technique. Moreover, dice score and c-Index were reported to evaluate the segmentation and prediction models, respectively. Result: For segmentation, the weighted fusion technique followed by the Cascade-Net segmentation algorithm resulted in a validation dice score of 72%. External testing of 71% confirmed our findings. DFs extracted from sole PET and MI followed by RSF enabled us to receive a validation c-index of 66% for RFS prediction. The external testing of 59% confirmed our finding. Conclusion: We demonstrated that using the fusion technique followed by an appropriate automated segmentation technique provides good performance. Moreover, employing DFs extracted from sole PET and HMLS, including MI linked with RSF, enables us to perform the appropriate survival prediction. We also showed imaging information extracted from PET outperformed the usage of the fused images in the prediction of RFS.