Prediction of Head-Neck Cancer Recurrence from Pet/CT Images with Havrda-Charvat Entropy

This paper proposes a loss function based on Havrda-Charvat entropy in deep neural networks for outcome prediction in head-neck cancers. Havrda-Charvat is a parameterized cross-entropy which generalizes the classical Shannon-based cross-entropy. Its parameter denoted α takes its values in ]0, ∞[ and one can recover some usual entropies, for instance Shannon for α = 1 or Gini coefficient for α = 2. In this paper, we propose to use this entropy to predict cancer recurrence by incorporating it in a neural network instead of Shannon's entropy for better adaptability. Our deep network is composed of a double auto-encoder to extract features and a classifier to predire cancer outcome. The experiments are conducted on MICCAI challenge dataset of Head-Neck cancer. The influence of the parameter on the results is studied and an optimal interval of its values is found. A result shows that Havrda-Charvat entropy can achieve better prediction performance than Shannon entropy, which is the most widely used in prediction task nowadays.