Back to Projects
In Production / Thesis Validation

MSX-Sentinel: Anatomically-Aware Diagnostic Framework

A multi-stage AI pipeline bridging high-performance computing (HPC) and Clinical VLMs to automate pathogen detection and host-response analysis in aquaculture.

MSX-Sentinel: Anatomically-Aware Diagnostic Framework

About the Project

Digital pathology for aquaculture faces a unique challenge: massive 130GB Whole Slide Images (WSI) paired with extremely rare clinical targets. I architected 'MSX-Sentinel,' a three-tier cascade system that mimics the workflow of a senior pathologist. The system utilizes a top-down approach: first, it performs macro-anatomical mapping to isolate individual specimens and identify key organs (Gills, Mantle, Gonads). Second, it deploys a high-speed YOLO-based 'Scout' to identify parasitic candidates. Finally, a Vision-Language Model (VLM) acts as the 'Clinical Eye,' performing structured morphological analysis on high-resolution patches. By integrating a quantitative biomarker layer (Hemocyte Density counting), the system provides an explainable diagnostic grade (0-3) that aligns with expert standards while reducing manual scanning time.

Engineering Challenges

01Multi-Scale Coordinate Synchronization

Mapping findings from a 1.25x macro-view down to a 40x high-resolution parasite location across a 130GB image file.

Solution: Developed a custom Level 0 Global Coordinate normalization system. All detections and anatomical regions are mapped to a unified pixel-grid, allowing for seamless 'drill-down' navigation between different magnification levels without data loss.

02Domain-Specific VLM Hallucinations

Generalist AI models lack the visual vocabulary for bivalve histology, often confusing host immune cells with parasites.

Solution: Implemented a 'Structural Ontology' prompting strategy. Instead of open-ended descriptions, the VLM is forced to fill a strict clinical JSON schema grounded by visual few-shot exemplars, increasing morphological verification accuracy.

Interested in this project?

View More Work