Date on Master's Thesis/Doctoral Dissertation

8-2025

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Electrical and Computer Engineering

Degree Program

Electrical Engineering, PhD

Committee Chair

Ali, Asem

Committee Member

Zurada, Jacek

Committee Member

Yampolskiy, Roman

Committee Member

McIntyre, Michael

Author's Keywords

Medical image segmentation; deep learning; few-shot

Abstract

This dissertation explores the modeling and analysis of medical images, focusing on the intricate task of colon segmentation and subsequent 3D reconstruction, which are critical steps in Computed Tomography Colonography (CTC) systems. The primary objective of this research is to develop precise segmentation approaches to enhance the accuracy of colon identification and reconstruction from abdominal CT scans. Three distinct segmentation approaches are proposed and evaluated: a Markov Random Field (MRF)-based approach, a convolutional neural network (CNN)-based deep learning (DL) approach, and a sequential episodic training with dual contrastive learning Approach (G-SET-DCL) that has a flavor of few-shot learning (FSL). To address the challenges of colon segmentation, an MRF-based method integrating image appearance and spatial interactions is subsequently applied on CT scan starting from the rectum to isolate the colon region and generates initial segmentation by framing the problem as a graph partitioning task. Recognizing the limitations of classical algorithms, particularly in handling complex colon shapes and noisy data, this dissertation proposes two novel DL approaches to boost the performance of colon segmentation techniques. The CNN-based DL approach enhances 2D segmentation models by incorporating 3D contextual information through an innovative attention map technique. This method captures crucial details for precise colon segmentation without the complexity of 3D CNNs or Long Short-Term Memory (LSTM) networks. This approach is validated on a private dataset of 98 CT scans from 49 patients (each patient having both supine and prone scans, totaling 42,609 2D images), demonstrating significant performance improvements with a validation Dice coefficient (DC) of 98.76% (Jaccard index, JD 97.60%) compared to the MRF-based approach's 87.90% (JD 84.50%). A major challenge in medical imaging is the scarcity of annotated data. To address this, we adopted the principles of Few-Shot Learning (FSL), aiming to leverage information from minimally labeled data to process (segment) more unlabeled data. This approach is inspired by the anatomical consistency of the human body, particularly in abdominal CT scans, where structures typically exhibit gradual changes across consecutive slices. Specifically, if a pixel is identified as part of the colon in one slice, it is highly likely to remain part of the colon in the next or previous few slices. Our method begins with rectum detection using an MRF approach, a region that can be easily and reliably identified. From this starting point, we sequentially segment the colon slice by slice, with each slice segmented based on information from its neighboring slice. Specifically, this approach leverages a sequential episodic training strategy that incorporates 3D contextual information. Consecutive CT slices are treated as Support and Query pairs, where the Query slice is segmented based on its preceding Support slice, effectively capturing anatomical smoothness and 3D context. Moreover, our method, G-SET-DCL, is designed to exploit the gradual anatomical transitions seen in consecutive CT slices, particularly in the colon. During training, each episode consists of two consecutive CT slices — a Support slice providing context and a Query slice for segmentation. This setup ensures that the model leverages spatial continuity between slices, enhancing segmentation coherence. To further enhance model performance, we incorporate negative samples — unrelated slices lacking the colon — which help the model distinguish colon features from irrelevant structures. As a result, these negative samples improve the model’s ability to differentiate between colon and non-colon regions, reducing false positives. Finally, by training with episodic segmentation, the model incrementally learns to handle diverse and challenging segmentation tasks, maintaining consistency across variations in the data and improving overall segmentation accuracy. Experiments on both a private dataset of 98 CT scans and a public dataset of 30 CT scans illustrate that the proposed FSS model achieves a remarkable validation DC of 97.30% (JD 94.50%) compared to classical FSS approaches' 82.10% (JD 70.30%). Our findings highlight the efficacy of sequential episodic training in accurate 3D medical imaging segmentation. This research underscores the effectiveness of integrating classical and deep learning techniques for precise and efficient colon segmentation and 3D reconstruction, significantly advancing the capabilities of CTC systems.

Download

Share

COinS