|
|
Using unlabeled data for retinal segmentation
Shemshur, Andrii ; Jakubíček, Roman (oponent) ; Vičar, Tomáš (vedoucí práce)
This bachelor's thesis is concerned with the development and evaluation of advanced methods for medical image segmentation in the context of limited training data. The study examines supervised learning techniques employing Convolutional Neural Networks (CNNs), transfer learning with pre-trained models, and semi-supervised learning strategies. A supervised convolutional neural network (CNN) model based on the U-Net architecture was employed as the baseline, achieving a Dice coefficient of 77.6\% and an intersection over union (IoU) of 63.4%. The application of transfer learning using a ResNet34 encoder pre-trained on ImageNet led to a notable improvement in performance, with a Dice coefficient of 81.9%, an IoU of 69.3%, and an accuracy of 96.7%. Furthermore, semi-supervised learning strategies, including pseudo-labeling and denoising pretraining, were employed to enhance the model's performance. The pseudo-labeling approach yielded a Dice coefficient of 81.7% and an IoU of 69.1%, thereby demonstrating the efficacy of leveraging unlabeled data. The denoising pretraining approach demonstrated robust performance, achieving a Dice coefficient of 80.3% and an IoU of 67.0%, even in the presence of noisy and unlabeled data. These outcomes underscore the potential of transfer learning and semi-supervised methods to enhance segmentation accuracy in medical image analysis. They provide a robust foundation for future research in this field.
|
host ::
RSS.