JMIR Publications

ABSTRACT

Background:

Cancer has become the second leading cause of death globally. Most cancer cases are due to genetic mutations, which in return affects the metabolism and further result in facial changes.

Objective:

In this study, we aim to reveal the facial features of cancer patients using deep learning techniques.

Methods:

8124 face images of cancer patients are collected to build the cancer face image dataset. And the non-cancer face image dataset is built by randomly selected from the public available MegaAge dataset according to the sex and age distribution of cancer face image dataset. The face image is pre-processed to obtain an upright centered face chip, of which the background is filtered out to exclude the effect of non-relative factors. A ResNet is constructed to classify cancer and non-cancer cases. Transfer learning and mini-batch, few epochs, L2 regulation and random dropout training strategies are used to prevent over-fitting. And guided gradient weighted class activation mapping (grad-CAM) is used to reveal the relevant features.

Results:

Both the male and female cancer average faces display more obvious facial adiposity than the non-cancer average faces, which was supported by the landmark comparison. The training process terminates after 5 epochs. The area under receiver operating characteristic (ROC) curve (AUC) is 0.94 and the accuracy rate is 0.82 on testing dataset. The relative features of cancer case are mainly the features of facial skins, while the relative features of non-cancer case are extracted from the complementary face region.

Conclusions:

In this study, we built up a face dataset of cancer patients, and constructed a deep learning model to classify cancer and non-cancer faces. We found that the facial skin and adiposity were closely related with cancer cases.