Superpixel labeling for medical image segmentation

Nowadays, most methods for image segmentation consider images in a pixel- wise manner, which is a huge job and also time-consuming. On the other hand, superpixel labeling can make the segmentation task easier in some aspects. First, superpixels carry more information than pixels because they usually follow the edges present in the image. Furthermore, superpixels have perceptual meaning, and finally, they can be very useful in computationally demanding problems, since by mapping pixels to superpixels we are reducing the complexity of the problem. In this thesis, we propose to do superpixel-wise labeling on two med- ical image datasets including ISIC Lesion Skin and Chest X-ray, then we feed them to the U-Net Convolutional Neural Network (CNN) DoubleU-Net and Dual-Aggregation Transformer (DuAT) network to segment our images in term of superpixels. Three different methods of labeling are used in this thesis: Su- perpixel labeling, Extended Superpixel Labeling (Distance-base Labeling), and Random Walk Superpixel labeling. The Superpixel labeled ground truths are used just for training. For the evaluation, we consider the original image and also the original binary ground truth. We considered four different superpixel algorithms, namely Simple Linear Iterative Clustering (SLIC), Felsenszwalb Hut- tenlocher (FH), QuickShift (QS) , and Superpixels Extracted via Energy-Driven Sampling (SEEDS). We evaluate the segmentation result with metrics such as Dice Coefficient, Precision, Intersection Over Union (IOU), and Sensitivity. Our results show the accuracy of 0.89 and 0.95 percent in dice coefficient for skin lesion and chest X-ray datasets respectively. Key Words: Superpixels, Medical Images, U-Net, DoubleU-Net, Image seg- mentation, CNN, DuAT, SEEDS.