Towards Clinical-Grade Pathology AI: From Whole-Slide Evidence Discovery to Clinical Translation and Decision Support

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Towards Clinical-Grade Pathology AI: From Whole-Slide Evidence 
Discovery to Clinical Translation and Decision Support"

By

Mr. Zhengrui GUO


Abstract:

Pathology is the cornerstone of modern cancer care, providing essential 
evidence for disease diagnosis, prognosis, and treatment planning. 
Whole-slide imaging turns glass slides into high-resolution digital forms 
that can be analyzed computationally, but clinical translation requires more 
than accurate prediction alone. Pathology AI must reason over gigapixel 
tissue context, learn from limited expert supervision, translate visual 
evidence into clinically interpretable descriptions, and be validated in 
settings that reflect real-world diagnostic workflows. This thesis develops a 
connected pathway towards clinical-grade pathology AI, moving from 
whole-slide evidence discovery to clinical translation and 
pathologist-centered decision support.

The thesis proceeds through four connected contributions. Querent addresses 
whole-slide evidence modeling by introducing query-aware dynamic 
long-sequence modeling, which estimates relevant tissue context for each 
query region and uses selective attention to make long-range WSI reasoning 
computationally feasible. FOCUS addresses learning under scarce annotated 
data through knowledge-enhanced adaptive visual compression, combining 
pathology foundation-model representations with language-derived diagnostic 
cues to reduce redundant visual content and prioritize diagnostically 
informative regions for few-shot WSI classification. HistGen moves from 
diagnostic prediction to evidence translation by linking local morphologic 
evidence, global WSI context, and pathology-report language through a 
local-global visual encoder and cross-modal context interaction. These 
directions are ultimately brought together in PulmoFoundation, a 
lung-specific pathology foundation model evaluated through retrospective 
internal and external validation, prospective validation, triage-oriented 
analysis, and a reader study of pathologist-AI interaction, shifting the 
focus from benchmark performance alone to whether foundation models can 
support triage, assist pathologists, and function reliably within real-world 
diagnostic workflows.

Together, these works support the thesis's central argument: clinical-grade 
pathology AI cannot be built from prediction accuracy alone. It requires 
coordinated progress in representation learning, data- and label-efficient 
learning, evidence translation, translational validation, and human-AI 
interaction. Rather than advocating autonomous replacement of pathologists, 
this thesis argues for validated, monitored, and human-centered AI systems 
that organize diagnostic evidence, communicate interpretable findings, 
support expert judgment, and make computational pathology more reliable for 
clinical decision support.


Date:                   Monday, 27 July 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Prof. Qian LIU (IEDA)

Committee Members:      Dr. Hao CHEN (Supervisor)
                        Prof. Gary CHAN
                        Dr. Yinghao XU
                        Dr. Terence Tsz Wai WONG (CBE)
                        Prof. Pong Chi YUEN (HKBU)