Click here to


Are you sure ?

Yes, do it No, cancel

A 3D Deep Convolutional Neural Network for Lung Cancer Survival Prediction Using Transfer Learning

M Ibrahim1, D Visvikis1 , C Cheze Le Rest1,2 , M Hatt1 (1) Laboratoire de Traitement de l'Information Medicale (LaTIM-UMR 1101 INSERM), IBSAM, UBO, UBL, Brest, France (2) Service de medecine nucleaire, CHU Miletrie, Poitiers, France


(Tuesday, 7/31/2018) 7:30 AM - 9:30 AM

Room: Davidson Ballroom B

Purpose: To design a 3D deep Convolutional neural network for lung cancer survival prediction using transfer learning.

Methods: Transfer learning can be useful in medical imaging where availability of large datasets are limited. Sufficient dataset size is a crucial requirement for efficient training of deep learning techniques.Efficient design of convolutional neural networks(CNN) is also important due to the high number of hyperparameters considered in these models. We investigated the use of a 3D Deep CNN to predict survival rate of lung cancer patients from computed tomography (CT) images using transfer learning.The process was a two step approach.First,we exploited a publicly released dataset of lung cancer patient’s diagnostic CT (n=1397) to train a CNN to differentiate normal from cancerous cases.Second,the same CNN was fine tuned for our own specific application:identifying patient with poor prognosis (≤6 months overall survival) in a smaller cohort (n=110) of non-small cell lung cancer (NSCLC) patients using the low-dose CT component of PET/CT acquisitions.In this second step,the initialization was carried out using the CNN built in the first step that had already learned on the larger database of the CT image features of cancerous lungs.The impact of some hyperparameters (e.g batch processing) were also evaluated.

Results: The trained CNN achieved 85% testing accuracy in distinguishing cancerous from normal CT scans in the larger database.After transfer learning,the CNN was able to reach 84.85% testing accuracy in identifying the NSCLC patients with ≤6 months survival.Batch processing choices had a significant impact on overall performance.In both steps,70% of the available dataset was used for training/validation and 30% for independent testing.

Conclusion: This work illustrates that the transfer learning proved efficient to overcome the limitation of small datasets usually available in oncology studies. The designed workflow could be further improved by thoroughly investigating the impact of choices of various hyperparameters.

Funding Support, Disclosures, and Conflict of Interest: Funding Acknowledgements : Brittany Region and Labex Cominlabs (Project RAMPART),France


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email