View Proposal

---

Proposer

Marta Vallejo

Title

Quantitative Explainable AI for Gut Digital Pathology Using Local Vision-Language Models and LayerCAM Analysis (in collaboration with the University of Aberdeen)

Goal

Develop an AI pipeline that uses local multimodal large language models (LLMs) to identify major tissue compartments in gut histology images and combines these segmentations with LayerCAM explainability maps to quantitatively analyse which tissue structures contribute most to deep learning-based disease classification.

Description

Deep learning models have shown excellent performance in digital pathology, but understanding why they make a particular prediction remains a major challenge. Explainability methods such as LayerCAM can highlight image regions that influence a model's decision, yet these visualisations are typically interpreted qualitatively rather than quantitatively. This project aims to develop an automated pipeline that identifies the main tissue compartments in gut histology images, including the lamina propria, crypts and surrounding tissue, using local multimodal large language models (LLMs) and modern computer vision techniques. The resulting tissue maps will then be combined with LayerCAM outputs from an existing deep learning classifier to measure how much model attention is allocated to each anatomical structure. The project will involve evaluating suitable open-source local AI models, designing the segmentation workflow, developing quantitative metrics for explainability analysis, and validating the results on a real research dataset related to neurodegenerative disease. The final outcome will be a reusable framework for quantitative explainable AI in digital pathology, with potential applications to other tissue types and disease areas, and the possibility of contributing to a scientific publication.

Resources

(1) A curated gut histology image dataset from ongoing neurodegenerative disease research. (2) Existing deep learning classification models and LayerCAM explainability pipeline developed within the ML-Health Research Group.

Background

This project combines computer vision, multimodal large language models (LLMs), digital pathology and explainable artificial intelligence (XAI). Students will gain experience in deep learning, image segmentation, foundation AI models, and quantitative interpretation of neural network decisions using real biomedical research data.

Url

Difficulty Level

Variable

Ethical Approval

Full

Number Of Students

2

Supervisor

Marta Vallejo

Keywords

digital pathology, explainable artificial intelligence, xai, large language models, multimodal ai, computer vision, image segmentation, deep learning, histopathology, gut tissue analysis, layercam, biomedical image analysis, neurodegenerative diseases

Degrees

Bachelor of Science in Computer Science
Master of Engineering in Software Engineering
Master of Science in Artificial Intelligence
Master of Science in Computing (2 Years)
Master of Science in Data Science
Master of Science in Robotics
Master of Science in Software Engineering
Bachelor of Science in Computing Science
Bachelor of Engineering in Robotics
Postgraduate Diploma in Artificial Intelligence