Bilkent University
Department of Computer Engineering


Fully Convolutional Networks for Semantic Segmentation of CT Scans Using Fractal Maps


Aziza Saber Jabdaragh
Master Student
(Supervisor: Assoc. Prof. Çiğdem Gündüz Demir )
Computer Engineering Department
Bilkent University

Abstract: In this talk, we present the potential for the information in fractal dimensions to be utilized properly in a multi-task network for the segmentation of images in CT scans. Over the recent years, fully convolutional networks such as U-NET and its variations have gained state-of-art results in most semantic segmentation tasks in medical image analysis. Due to the fact that such single-task networks cannot perform well when the dataset has varying shapes and overlapping boundaries, multi-task networks have been introduced to segment the images in such hardto- segment areas. Most of these networks use the shape and contour information of the objects as an auxiliary task to better learn the main task of segmentation. However, despite the successful results of such networks in most cases, thesenetworks suffer from considerable limitations on the images that the objects are fuzzy, fragmented, and rugged where their structure is complex and does not follow the smoothness of objects in Euclidean distance space. Moreover, objects might have very irregular and noncontinuous boundaries and thus they cannot be accurately measured using the normal dimensions to define their structure through the shape or similar information while their shapes can vary widely making the shape aware networks struggle to learn all types of the shapes. To the best of our knowledge, such problems are not addressed properly, and thus, to alleviate these problems, we present a novel end-to-end framework that utilizes the fractal dimensions through a multi-task network on the segmentation of CT scans. More specifically, we model the complexity and detail of pattern change in the object structure when changing the scale, and overall the roughness of objects that is associated with the challenges that the other methods do not consider during the modeling process. The proposed model is evaluated on the Left Atrium segmentation which is required for the evaluation of many cardiovascular diseases.


DATE: 11 April 2022, Monday @ 15:30 Zoom