Several research papers have been published over the peroid of the PhD journey:
- GoogleScholar publication profile.
- Total of 8+ publications. Future publications "under review".
- h-index 4 and 114 Total Citation since 2020.
- Research domain is about Deep Learning models applied to Colonoscopy to identify abnormal regions.
Type : PhD Thesis
Year : 2024
Publisher: University of Nottingham
Summary
This PhD thesis develops deep learning models for real-world clinical colonoscopy, focusing on bowel preparation assessment and polyp segmentation. It introduces novel approaches to improve model generalizability and addresses limitations in previous works, arguing that reported results are exaggerated due to inappropriate dataset setups.
Selected Figures
Type : Conference Paper (BICS)
Year : 2023
Publisher: Springer
Summary
The paper proposes methods to improve deep learning models for polyp segmentation in colonoscopy images. By using online spatial interpolation and hue transformation, the approach enhances model robustness and generalizability across different datasets and clinical conditions, improving segmentation accuracy in varied environments.
Selected Figures
Type : Conference Paper (ISBI)
Year : 2023
Publisher: IEEEXplore
Summary
This paper proposes a method to improve the generalizability of deep learning models for polyp segmentation in colonoscopy images. The authors introduce a technique that minimizes the total variation of images during training, which helps reduce noise and preserve important structural details. This approach enhances the model's ability to generalize across diverse datasets and imaging conditions, leading to more accurate and consistent polyp segmentation in real-world clinical scenarios.
Selected Figures
Type : Conference Paper (AIHA@ICPR)
Year : 2022
Publisher: Springer
Summary
The paper presents a deep learning-based approach for automating the evaluation of bowel preparation in colonoscopy images. The authors develop a model that can accurately assess the quality of bowel preparation by analyzing colonoscopy video frames, enabling faster and more objective assessments compared to traditional methods. This approach aims to improve the efficiency and consistency of bowel preparation evaluations in clinical practice.
Selected Figures
Type : Conference Paper (MedAI)
Year : 2021
Publisher: Nordic Machine Intellegince
Summary
The paper proposes an enhancement to the DeeplabV3 segmentation model by incorporating a Gated Recurrent Unit (GRU) to combine feature maps more effectively. The authors argue that by using GRU, the model can capture better contextual information across different layers, leading to improved segmentation accuracy. This approach aims to refine DeeplabV3’s performance, particularly in tasks requiring detailed and precise segmentation.
Selected Figures
Type : Journal Paper (????)
Year : 2025
Publisher: ????
Summary
The paper (under-review) evaluates the current state of deep learning models for bowel preparation assessment in colonoscopy. The authors discuss the progress made in automating this task, highlighting the strengths and limitations of existing approaches. They argue that while deep learning has shown promise, challenges remain in achieving consistent, accurate results across diverse datasets and clinical settings, suggesting that the problem is not yet fully solved.
Selected Figures
Type : Journal Paper
Year : 2025
Publisher: ????
Summary
The paper (writing stage) introduces a novel approach for improving the generalizability of polyp segmentation models by optimizing a nested negative-log likelihood (N-NLL) loss function. This method aims to address challenges like overfitting and dataset variability, enhancing the model's performance across different clinical settings and data sources. The optimization of the nested N-NLL helps create more robust and adaptable segmentation models for real-world applications.
Selected Figures
