Automatic Detection of Infantile Hemangioma using Convolutional Neural Network Approach

Infantile hemangioma is the most common tumor of childhood. This study proposes an automatic detection as a preliminary step for a further accurate monitoring tool to evaluate the clinical status of hemangioma. For the detection of hemangioma pixels, a convolutional neural network (CNN) was trained on patches of two classes (hemangioma and nonhemangioma) from the train dataset, and then it was used to classify all the pixels of the region of interest from the test dataset. In order to evaluate the results of segmentation obtained with CNN, the region of interest of the test dataset was also segmented using two classical methods of segmentation: fuzzy c-means clustering (FCM) and segmentation on histogram using the Otsu method. Given the relatively small dataset and high variance of clinical appearance, the CNN model reached 91.02% global score on the validation set becoming a promising tool for hemangioma evaluation.