Figure 2
From: Semi-supervised segmentation of retinoblastoma tumors in fundus images
(a) Sample of results regarding segmentation of tumors, using clustering. Given the nature of this method, the achieved boundaries are expected to be as precise as possible. Moreover, using more powerful hardware capable of processing fundus images with higher resolution, smaller signs of retinoblastoma can be detected. However, when compared to the suggested supervised method, the computational cost of this approach is significantly higher. (b) Successful Results of segmentation refinement based on the first approach, where the model is trained using original images as input data and manually edited outputs of unsupervised method as output data. Note that value of each pixel in final results is mapped to one or zero, using a threshold of 0.5. The suggested method has successfully achieved its primary objective for the majority of candidate images. However, it should be noted that this method is not foolproof, which may be due to the small size of the network or the low resolution of input images. Further research and development may be required to overcome these limitations and improve the reliability of the method.
