Bios: Luqmaan Shaikh

Program Track: Skills Development

GitHub Username:

What was your favorite seminar? Why?

If you were to summarize your summer internship experience in one sentence, what would it be?

Blog Post

Abstract:

The NASH Fibrosis Staging: Virtual Staining and Deep Ordinal Regression project aims to revolutionize the common assessment of liver fibrosis by integrating advanced digital image analysis with clinical serology data. This project focuses on the application of virtual staining techniques and deep ordinal regression models to precisely quantify fibrotic scarring and identify key biomarkers associated with non-alcoholic steatohepatitis (NASH) progression. By offering a more accurate and detailed staging of fibrosis, the project enhances the early detection and classification of chronic liver diseases, which are often marked by fibrosis. The predictive model developed will facilitate early diagnosis and intervention, ultimately improving patient outcomes. Given that liver diseases account for over 2 million deaths globally each year, according to the World Health Organization (WHO), this project addresses a critical need for more effective and timely approaches to liver disease management.

Methods:

To develop the predictive model, whole slide images (WSI) of liver fibrosis were obtained from Dartmouth Health and meticulously labeled by pathologists. Given the extensive size of these images, they were divided into smaller patches for more manageable processing. The first step involved distinguishing liver tissue from the background by applying an RGB filter, which allowed for the separation of the relatively white background from the liver tissue, setting the background to pure white for easy identification. This was essentially to create virtual images for training, although we were also given labeled WSI images for training as well.

An Ordinal Regression Graph Convolutional Network (GCN) model was then trained, though at this stage, it did not yet incorporate serology data. The model's accuracy in classifying fibrosis stages was limited, likely due to the absence of serology data, the variability in whole slide images, and occasional disagreements in pathology classifications. Despite these challenges, the model continued to train based on epoch data graphs.

The project reached this developmental stage within the given timeframe. Future steps include Optimizing and refining the GCN model architecture. Integrating serology data more effectively with imaging data, Incorporating statistical analysis of pathologist discrepancies to improve model accuracy, and Exploring additional deep learning architectures and advanced techniques

While significant progress was made, further work is needed to achieve the project's ultimate goal of creating a reliable model for predicting liver fibrosis presence and severity.

Conclusion:

Our initial efforts to use a (GCN) for staging liver fibrosis in NASH patients show promise, particularly in identifying broader trends within the data. However, the model currently struggles with accurately distinguishing intermediate fibrosis stages, as evident from our results. These challenges highlight the need for further optimization and development. By addressing these areas, we aim to develop a more robust and reliable model for NASH fibrosis staging, ultimately improving diagnostic accuracy and patient outcomes in liver disease care. It was challenging, interesting, and fun to use machine learning algorithms and learn about their effects on pathology as part of my first real research experience, which I believe turned out to be an enriching and beneficial experience.