Show abstract - Show schedule - Proceedings - PDF - Reviews

Introduction: Prognostic models for cancer patients may improve decision making to personalize management of cancer patients. In the current study, we propose a deep learning-based predictive model for Non–Small-Cell Lung Cancer patients. We developed a combined 2D and 3D model to include all 3D tumour information in CT images without losing spatial information. Method: In the current study, we enrolled 363 histopathological proven Patients from 5 different centres gathered from TCIA. In current study we aimed to predict 2 year overall survival, for this purpose continues value of survival times (calculated from start of the treatment) were dichotomized by 2-year cut-off point. We excluded right-censored patients (alive patients with follow up less than 2 years). We extracted each tumour separately based on bounding box We applied largest bounding box on each tumour separately considering the centre of mass of each lesion. Finally, we extracted a tumour region with size of 128×128×Z which Z is different for each different patient. We implemented 2D networks to construct 3D combination CNNs network. Feature extraction was performed in each 2D slices separately using same architecture and then final layer of network were averaged to train in 3D volume of tumor. Different architecture of networks including simple Xception, VGG, ResNet, Inception, and DensNet were implemented in this approach. Data of three centre (257) were used for train/validation and two centres were hold for external test set (106 patients). The predictive power of each model was assessed by using the area under the receiver operating characteristic (AUC - ROC), precision, recall, and accuracy. DeLong tests were performed on AUC to compare different models. Results: We set up a study to predict two year overall survival in NSCLC patients using deep neural networks. In term of accuracy, VGG had highest performance of 0.75. In term of precision and recall Xception and VGG had highest value of 0.77 and 0.83, respectively. Considering all four parameters, VGG had highest performance with precision, recall, AUC, and accuracy of 0.72, 0.83, 0.78, and 0.75, respectively . There was no statistically significant difference between different models (delong test p-value >0.05). Conclusion:. In current study we proposed end-to-end prognostic modelling in NSCLC using deep neural networks and pre-treatment CT images. Notwithstanding high variability across different datasets including geographic, ethnicity and CT scanner, image acquisition and image reconstruction, proposed models performed very well on different centres.
Hide abstract

Friday 8th July
Poster Session 3.2 - onsite 11:00 - 12:00, virtual 15:20 - 16:20 (UTC+2)
Hide schedule