Deep learning based brain tumor segmentation in MRI images
Loading...
Date
2025-02-17
Authors
Τσουμπού, Ισμήνη
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Brain tumor segmentation is a critical task in medical image analysis, playing a vital role in diagnosis, treatment planning, and patient monitoring. Despite significant progress in the field, the complexity of brain tumor morphology and variability in MRI data pose significant challenges. In this thesis, we develop and evaluate two deep learning-based models for brain tumor segmentation using 3D MRI volumes: a hybrid 3D model that combines U-Net and transformer architectures, and a similar computationally efficient 2D slice-by-slice model that includes an initial slice-level clinician’s prediction.
Both models were evaluated on a well-established brain tumor dataset, achieving performance comparable to state-of-the-art methods. The 3D model delivered competitive results; however, its high computational demands required downsampled images during training, resulting in a trade-off in accuracy when applied to full-resolution data. In contrast, the 2D slice-by-slice model emerged as a more lightweight and computationally efficient solution. By sequentially processing each MRI slice alongside the previously predicted mask slice, it effectively preserves critical image details without downsampling and maintains, to some extent, spatial coherence across slices. This strategic design enables accurate segmentation while mitigating the computational challenges posed by high-resolution data.
This study highlights the potential of hybrid CNN-transformer models for medical image segmentation, providing valuable insights into their design, performance trade-offs, and practical applicability. The findings contribute to advancing brain tumor segmentation methodologies, paving the way for further research and optimization in medical imaging tasks.
Description
Keywords
Artificial intelligence, Deep learning, Brain tumor segmentation