Advancing computational pathology: generation, generalization, evaluation, and automation

PhD Thesis Proposal Defence


Title: "Advancing computational pathology: generation, generalization,
evaluation, and automation"

by

Mr. Jiabo MA


Abstract:

Histopathological examination is the gold standard for cancer diagnosis. The
digitization of glass slides into whole-slide images has shifted pathology from
qualitative microscopy toward large-scale quantitative analysis. Recent advances
in artificial intelligence have further expanded this potential by linking
tissue morphology to molecular alterations and clinical outcomes. Consequently,
digitized tissue can now be used not only to assist diagnostic assessment, but
also to investigate complex biological and clinical questions at scale. However,
Realizing this potential requires addressing three practical challenges: how
much biological signal can be extracted from routine images, whether learned
representations remain reliable across diverse tasks and patient populations,
and how easily domain researchers can apply these computational tools in
practice. This thesis develops computational frameworks that bridge the gap
between raw tissue images and actionable biological and clinical insights.
Specifically, virtual staining offers a cost-effective way to extract additional
molecular and structural information from existing images without performing
physical staining in the laboratory. Training these models typically relies on
paired microscopy data, where corresponding images often suffer from spatial
misalignment. To improve the quality of virtual stained images, this thesis
introduces a supervised pretraining strategy and a registration strategy.
Evaluated through quantitative image quality metrics, downstream diagnostic
tasks, and blinded pathologist assessments, these methods substantially improve
the fidelity and robustness of virtual staining. Beyond synthesizing additional
image channels, supporting diverse biological and clinical inquiries requires
generalizable feature representations. Existing models trained with different
datasets and objectives capture distinct aspects of tissue morphology, causing
their performance to vary across downstream tasks. To integrate these
complementary strengths, this thesis presents the Generalizable Pathology
Foundation Model, which combines self-distillation with knowledge distillation
from multiple expert encoders. Pretrained on 190 million patches from 72,280
whole-slide images, the model ranks first on 42 of 72 evaluation tasks,
demonstrating that unifying diverse representations broadens the range of tasks
a single model can support. The practical value of foundation models ultimately
depends on whether their representations remain reliable in real-world clinical
settings. This thesis therefore establishes PathBench, a large-scale clinical
benchmark comprising 84,944 whole-slide images from 41,177 patients across 26
hospitals. For domain researchers, the need to select and validate appropriate
models adds to the technical burden of data preparation, pipeline
implementation, and result interpretation. To lower this barrier, this thesis
introduces PathLab, an agentic system that translates natural-language research
objectives into executable analytical workflows. PathLab automatically selects
and integrates pathology models, generates code, adapts subsequent steps based
on intermediate outputs, and incorporates methodological checks throughout the
process. A user study with 33 participants demonstrates that pathologists and
medical students using PathLab can achieve predictive performance comparable to
that of senior AI researchers. Together, these studies connect algorithm
development with the practical requirements of pathology research. They
demonstrate how to recover richer information from unaligned tissue images, how
to unify complementary representations into a versatile foundation model, and
why rigorous clinical benchmarking is necessary for model deployment. By further
enabling automated analysis through natural-language interaction, this thesis
provides an end-to-end methodological foundation for using digitized tissue to
answer biological and clinical questions.


Date:                   Tuesday, 17 November 2026

Time:                   3:00pm - 5:00pm

Venue:                  Room 5501
                        Lifts 25/26

Committee Members:      Dr. Hao Chen (Supervisor)
                        Dr. Sehi L'Yi (Chairperson)
                        Dr. Terence Tsz Wai Wong (CBE)