Semi-Supervised Segmentation of Retinoblastoma Tumors in Fundus Images

Abstract

Retinoblastoma is a rare type of cancer that can appear in young children as the most common primary intraocular malignancy. Studies in developed and some developing countries have shown that more than 90% of children with retinoblastoma have been successfully cured thanks to early detections. The most common presenting sign is an unusual white reflection in the pupil. Medical experts may choose different approaches and treatments for retinoblastoma, based on the size, shape and location of the tumors. Given the high dependence of the process on prior knowledge, the results may vary. This study aims to present a model based on semi-supervised machine learning, with segmentation results comparable to the labeling done by medical experts. To do so, the gaussian mixture model is utilized to detect abnormalities in nearly 4200 fundus images. Due to the high calculational cost of this process, the results of this approach are used to train a cost-effective model for the same purpose. The second method demonstrated excellent results in the case of extracting highly detailed boundaries in fundus images. Using Sørensen–Dice coefficient as the comparison metric for the tasks of segmentation, an average accuracy of 77.19% on evaluation data has been achieved.