Session Item

Prognostic outcome models using clinical and image data make it possible to select the most optimal treatment method for individual oropharyngeal squamous cell carcinoma (OPSCC) patients. Deep learning based image feature extraction methods can identify more complex patterns. The aim is to build Cox models with the capability of predicting outcomes prior to treatment using clinical data, image features of CT extracted by an Autoencoder and the combination of clinical and image features and to compare their performance.

The OPC-Radiomics set (https://www.cancerimagingarchive.net/collections/) includes 606 oropharyngeal squamous cell carcinoma (OPSCC) patients. Pre-treatment planning CT, GTV contours of primary tumours (GTVp) and clinical parameters were available for 524 patients. The candidate clinical parameters for building outcome prediction models are gender, age, WHO performance status, TNM-STAGE, oropharynx cancer stage, treatment modality and HPV status. The endpoints were local recurrence (LR), regional recurrence (RR), local and regional recurrence (LRR), distant metastases free survival (MET), tumour-specific survival (TSS), overall survival (OS) and disease-free survival (DFS). The clinical parameters were analysed as categorical variables, except that age is normalized by dividing by 100. Tumour images of 64x64x64 pixels were extracted around the centre of mass of the GTVp contour. All pixel values outside the GTVp contour were set to zero.

As illustrated in Figure 1, we propose a deep learning based pipeline to extract image features for outcome prediction. Firstly, we utilize a 3D Autoencoder for extracting 1024 latent features that are representative for the tumour images of the cohort. The Autoencoder architecture is a pyramid Autoencoder for combining different-level features. The Adam optimizer with initial learning rate 0.001 was used to train the Autoencoders for 80 epochs. Secondly, for each outcome, a feed forward feature selection method was applied to select the most predictive features for the outcome prediction from both extracted image features and clinical features. Finally, the selected features are used to build Cox models that calculated the risk score on each outcome for individual patients. 

Table 1 displays the C-index results of the Cox models for each outcome on the training set and test set. The selected clinical features and the number of selected image features are also shown here. The Cox models built by clinical and image features together resulted in higher training C-index values than the models using only clinical or image features in all seven outcome predictions and higher testing C-index values in six outcome endpoints except for MET.      

Our proposed self-supervised learning based method which extracts image features of OPSCC patients and uses them in combination with clinical data shows better predictive performance than the clinical models for all outcome endpoints.