Unit:
Κατεύθυνση Βιοπληροφορική-Υπολογιστική ΒιολογίαLibrary of the School of Science
Author:
Tsagkalidou Athina
Supervisors info:
Μιχαήλ Φιλιππάκης , Καθηγητής, Τμήμα Ψηφιακών Συστημάτων, Πανεπιστήμιο Πειραιώς, (Επιβλέπων),
Ιωάννης Τρουγκάκος Καθηγητής Τμήμα Βιολογίας ΕΚΠΑ,
Βασιλική Οικονομίδου Αναπληρώτρια Καθηγήτρια Τμήμα Βιολογίας ΕΚΠΑ
Original Title:
Ανίχνευση και Ταξινόμηση όγκων εγκεφάλου με χρήση Deep Learning
Translated title:
Brain Tumor Detection and Classification Using Deep Learning
Summary:
Brain tumors represent a serious health challenge as they involve the uncontrolled growth of cellular mass in the brain. They can cause severe implications on the function of the nervous system and the patient's health. Early and accurate detection of brain tumors is crucial for the prevention and management of potential consequences. Detection methods include radiological examinations such as magnetic resonance imaging and the use of X-rays, as well as neurological examinations that may involve assessing neurological functions and evaluating symptoms. Diagnosis combined with appropriate therapeutic intervention can help manage brain tumors and improve the quality of life for patients.
Automated detection of brain tumors from magnetic resonance imaging (MRI) images is crucial for improving diagnosis and treatment of patients and can be achieved using advanced machine learning techniques. Machine learning algorithms, such as neural networks, can be trained for this purpose using MRI image data with relevant labels for the presence of brain tumors. This allows for automatic detection and potential classification of tumors with accuracy and speed, helping doctors to more effectively and quickly identify potential brain problems. The use of such advanced machine learning techniques can lead to more accurate diagnoses and improve patient care for those with brain tumors.
Within the scope of this thesis, two methods are presented and compared for both the problem of binary classification of the presence or absence of tumors in brains and for the problem of precise localization of brain tumors. These systems take MRI brain images as input. For the first problem, a performance comparison is made between a conventional convolutional neural network and a pre-trained ResNet50 model leveraging transfer learning and underwent model fine-tuning for problem-solving. The pre-trained classification model based on ResNet50 performed better as it achieved 98.4% accuracy as opposed to CNN which had 95.6% accuracy. More specifically, it successfully classified 1475 out of 1500 images (98.33%) without brain tumor, while 1477 out of 1500 images (98.47%) with brain tumor were correctly predicted. For the second problem, a performance comparison is made between two U-Net architectures, the main difference is the use of residual blocks. The method using the U-Net network and the residual blocks showed an accuracy of 95.15%, surpassing the method using the U-Net which had an accuracy of 94.14%. Using the U-Net network and the residual blocks the resulting error was 0.11, which is better compared to the U-Net method where it had an error equal to 0.35. The results showed that both problems can be effectively solved by the proposed methods, which have very small deviations in their predictions.
Additionally, a UI application was developed where the user can input MRI brain images and automatically receive predictions from the best-pretrained models developed for each respective problem.
Main subject category:
Science
Other subject categories:
Health Sciences
Keywords:
Artificial intelligence in radiology, oncology, brain cancer, brain tumor, machine learning, deep learning, neural networks, ResNet50, U-Net
