Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


SatyakiRoyChowdhury_MSThesis.pdf (9.77 MB)

ETD Abstract Container

Abstract Header

Receptive Field Expansion and Uncertainty-Guided Deep Learning for Adult Glioma Segmentation

Chowdhury, Satyaki Roy
http://orcid.org/0009-0006-7035-3743
http://rave.ohiolink.edu/etdc/view?acc_num=osu1744927089137318

Abstract Details

2025, Master of Science, Ohio State University, Electrical and Computer Engineering.
Accurate segmentation of brain tumors, particularly adult gliomas, is critical for effective diagnosis, treatment planning, and prognostication. Gliomas exhibit signifi- cant spatial and scale heterogeneity along with an infiltrative growth pattern, posing major challenges to conventional segmentation techniques. Conventional methods in brain tumor segmentation often fall short in capturing both fine local details and the broader contextual information needed for precise delineation of tumor bound- aries. Additionally, many current deep learning approaches lack robust uncertainty estimation, which is essential for reliably assessing prediction confidence in clinical decision-making. There remains a critical need for resource-efficient architectures that not only achieve high segmentation accuracy but also reduce computational costs, especially when addressing the complex, heterogeneous, and infiltrative characteristics of gliomas. In this thesis, we propose novel deep learning frameworks that have efficient receptive field expansion and uncertainty-guided segmentation to address these challenges. Our approach combines multi-scale attention mechanisms with Atrous Spatial Pyramid Pooling (ASPP) to capture both fine-grained local details and broader contextual information, thereby enhancing segmentation accuracy. Additionally, by incorporating Monte Carlo dropout and designing an uncertainty-aware loss function, our model quantifies per-pixel prediction confidence, ultimately improving the reliability and interpretability of the segmentation outcomes. Extensive experiments on the BraTS2023 and BraTS2024 datasets demonstrate that the proposed method outperforms traditional U-Net variants in terms of both segmentation performance and computational efficiency. The integration of uncertainty estimation not only offers valuable insights for clinical decision-making but also paves the way for more personalized treatment strategies in managing adult glioma. Following this, we introduce a comprehensive methodology to analyze and interpret the regions of interest (ROIs) that drive the predictions of deep neural networks in brain MRI segmentation. Our approach systematically investigates not only where the network focuses its attention but also which classical image features underlie its decision-making process. Specifically, we employ region-wise integrated occlusion technique to perturb different areas of the input image and quantify their influence on the segmentation outcome. Concurrently, we extract well established classical features—such as Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) that capture important textural, shape and intensity information. By correlating these classical features with the network’s internal activations, we gain valuable insights into the rationale behind a deep learning model’s predictions.
Irem Eryilmaz (Committee Member)
Golrokh Mirzaei (Advisor)
89 p.
Computer Science; Electrical Engineering
Atrous Spatial Pyramid Pooling; Uncertainty Guided Deep Learning; Interpretability; Computer Vision; Medical Image Segmentation

Recommended Citations

Citations

  • Chowdhury, S. R. (2025). Receptive Field Expansion and Uncertainty-Guided Deep Learning for Adult Glioma Segmentation [Master's thesis, Ohio State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=osu1744927089137318

    APA Style (7th edition)

  • Chowdhury, Satyaki Roy. Receptive Field Expansion and Uncertainty-Guided Deep Learning for Adult Glioma Segmentation. 2025. Ohio State University, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=osu1744927089137318.

    MLA Style (8th edition)

  • Chowdhury, Satyaki Roy. "Receptive Field Expansion and Uncertainty-Guided Deep Learning for Adult Glioma Segmentation." Master's thesis, Ohio State University, 2025. http://rave.ohiolink.edu/etdc/view?acc_num=osu1744927089137318

    Chicago Manual of Style (17th edition)

osu1744927089137318
26
© 2025, all rights reserved.
This open access ETD is published by The Ohio State University and OhioLINK.
Release 3.2.12