Biomedical Image Segmentation and Classification Using Deep Learning
Date Issued
2016
Date
2016
Author(s)
Gamero, Jonathan Ponce
Abstract
Processing of biomedical images is one of the most important tasks that medical institutions such as hospitals and research centers deal with in a day to day basis, but even though this is such an important and basic task technology and current approaches have never been able to be successfully used in practical situations mainly because of their low accuracy, this has started to change in the beginning of 2012 with the use of convolutional neural networks and deep learning to accurately classify and segment a variety of different images, including biomedical images. Although these new technologies are being widely used with incredible results in many diverse fields, their adoption in the medical community has been slow to say the least, this has been due to many different reasons like the fact that the training data needed to provide a good accuracy model needs to come from a wide variety of patients, hundreds at least, from all around the world, from all age groups and sexes. This in it by itself is a great challenge to put together for a medical institution let alone for an individual, and even after gathering the data, a great deal of effort needs to be put into anonymizing the data. This is extremely important because by law any medical records have to be completely private and cannot be used without the expressive permission of the patient, so in order to facilitate this data many hospitals will anonymize the data so it cannot be tracked down to the patient. The following thesis will try to provide an approach on segmenting and classifying biomedical images, specifically heart MRI images taken by a trained professional, using a series of steps, which include deep learning, to accurately determine the volume of the heart’s left ventricle when it is in diastole, the largest volume of a heart cycle, and systole, the smallest volume of a heart cycle, with this measurement one can calculate the ejection fraction, often described as the most important measurement for early detection of heart disease. As stated before an accuracy high enough to be used in actual medical scenarios will tried to be achieved. The data used for the training for the deep learning model was supplied by Kaggle, an online competition oriented platform designed to solve many of today’s difficult problems. What was also tried to achieve with this research was to create a way to track foreign objects such as cancer and different other diseases, first tracking the object of interest and then using deep learning to either classify, measure or detect the objects main characteristics.
Subjects
Deep Learning
left ventricle
ejection fraction
biomedical images
SDGs
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-105-R03922147-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):e5171e9968b1e4afef54b0a5a7880859