More about HKUST
Deep Learning for Computational Cytopathology: From Cellular Analysis to Whole-Slide Screening
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Deep Learning for Computational Cytopathology: From Cellular
Analysis to Whole-Slide Screening"
By
Mr. Hao JIANG
Abstract:
Computational cytopathology enables precancerous screening through
cytomorphological assessment of cellular specimens, yet clinical translation
faces fundamental challenges from morphological ambiguity, spatial crowding,
and domain shifts across multi-center cohorts. This thesis establishes a
systematic framework progressing from cellular analysis to whole-slide
screening. It first addresses categorical ambiguity and imbalance via HERO, a
holistic and historical instance comparison framework that enforces RoI-level
and class-level contrastive consistency with a confidence-gated memory bank,
enabling robust instance characterization under subtle inter-class
discrepancies and long-tailed distributions. For segmentation under spatial
overlapping, DoNet introduces a decomposition-recombination strategy with a
dual-path region segmentation module and semantic consistency-guided
refinement to disentangle translucent overlapping cell clusters, further
reinforced by a mask-guided region proposal module that imposes the biological
prior of nuclei-cytoplasm containment. To generalize structure-aware
refinement beyond cytology, GAInS develops anomaly-guided supervision through
gradient-field modeling and adaptive local refinement, precisely delineating
crossing, touching, and overlapping instances across diverse pathological
scenarios. At the patient level, these insights inform Smart-CCS, a
generalizable whole-slide screening architecture integrating self-supervised
pretraining on 127,471 whole-slide images from 48 centers, multi-instance
aggregation, and test-time domain adaptation with prototype alignment,
translating cellular evidence into actionable decisions across heterogeneous
clinical settings with prospective and histological validation. Together,
these methods establish an evolutionary path from cellular morphology to
holistic diagnostic reasoning, demonstrating that morphology-aware learning,
structure-aware segmentation, and generalizable aggregation can be integrated
to advance AI-assisted cytopathology for precancerous screening and patient
triage.
Date: Tuesday, 25 August 2026
Time: 3:00pm - 5:00pm
Venue: Room 5501
Lifts 25/26
Chairman: Prof. Jianan QU (ECE)
Committee Members: Dr. Hao CHEN (Supervisor)
Prof. Pedro SANDER
Dr. Dan XU
Dr. Terence Tsz Wai WONG (CBE)
Prof. Harry Jing QIN (PolyU)