Deep Learning Based Radiomics to Predict Treatment Response Using Multi-datasets

Abstract

In this work, we present a multitask network with multi datasets to assess the relapse of patients with head-neck and lung cancers after a therapy from both scanner images and patient clinical data. A multitask architecture is developed to realize classification of the multi-type of cancers and relapse prediction tasks using clinical data and radiomics features. Medical imaging requires reliable algorithms for analysis and processing, especially regarding diagnosis and outcome prediction. However, in medical domain, only small datasets are available, this is why we propose to combine several small data sets that contain the same type of images and patient information. We also propose to use Havrda-Charvat cross-entropy, which is a generalized cross-entropy with a parameter $\alpha$, as loss function for our traning step. It tends toward Shannon cross-entropy when said parameter $\alpha$ is equal to 1. The influence of the variations of the parameter on classification is assessed. The experiments are carried out on a dataset of 580 patients with two cancer datasets (head-neck or lung). The results assess that Havrda-Charvat entropy has slightly better performances in term of prediction accuracy: $64\%$ of correct prediction for Shannon’s entropy and at best $69\%$ of correct prediction for Havrda-Charvat for $\alpha =0.2$. The challenge is to find a suitable value of $\alpha$.