About
Sep 2026
I am an incoming Assistant Professor at Duke University. My research is machine learning for medical imaging: segmentation and detection under annotation budgets a clinical department can actually afford, models that pair volumetric images with the reports radiologists write, and evaluation that holds up across scanners and sites rather than on a single held-out split. What keeps me in this setting is that the constraints are concrete — annotation costs clinician hours, the text accompanying an image is a report rather than a caption, and a wrong answer has a patient attached to it.
My published work builds the methods this rests on: uncertainty-guided co-training for semi-supervised segmentation (CVPR 2022), novel-object captioning and referring-expression grounding (ICCV 2023, ACM MM 2023), 3D instance segmentation from click-level annotations, in-context semi-supervised learning, and generative models for reconstructing tissue structure (bioRxiv 2025).
Research
Medical imaging · three threads
Thread 01 · Annotation cost
Segmentation when the only qualified annotator is a clinician
A dense 3D annotation of a CT or MR volume costs radiologist hours, so the useful question is never how a model performs under full supervision — it is how much of the label can be dropped before the model stops being trustworthy. My CVPR 2022 work used disagreement between two prediction heads as an uncertainty signal for pseudo-labelling; ClickSeg pushed the annotation budget for 3D instances down to single clicks. I am carrying both into volumetric clinical data: organ and lesion delineation, longitudinal follow-up, and the common case where the positive class is a handful of voxels.
- Methods
- Co-training with uncertainty estimates, pseudo-label filtering, click- and scribble-level supervision, in-context adaptation to a new site
- Data
- CT / MR / PET volumes; DICOM → NIfTI pipelines that keep acquisition provenance
Thread 02 · Reports, not captions
Findings that point back at the voxels that produced them
A radiology report is not a caption. It is a set of claims, and each one should be traceable to evidence in the image — usually to a number a reader can check: a diameter, a volume, a Hounsfield-unit range, an SUV. My work on novel-object captioning (ICCV 2023) and referring-expression comprehension (ACM MM 2023) was that alignment problem in natural images; the clinical version adds calibration and auditing, which findings a model invents and which it quietly misses.
- Methods
- Contrastive alignment, anchor-based grounding, image–report pretraining, calibrated uncertainty
- Metrics
- Per-finding sensitivity at fixed specificity, grounded-report accuracy, calibration error
Thread 03 · Generalization
Performance that survives a change of scanner
Most imaging models degrade on the first scanner they were not trained on. I want to characterize that degradation across vendors, protocols, sites and patient subgroups, and to adapt a model to a new site with a realistic label budget instead of retraining it. A related thread is generative modeling of biological structure — reassembling tissue organization from dissociated measurements (bioRxiv 2025).
- Evaluation
- Dice/DSC for localized structures, AUROC / AUPRC with 95% confidence intervals, subgroup tables, external cohorts reported separately from the development set
Publications
Full list on Scholar
2025
In-Context Semi-Supervised Learning
Jiashuo Fan, Petru Rosu, Aaron T. Wang, Michael Li, Lawrence Carin, Xiang Cheng
arXiv:2512.15934 · Preprint
2025
Model Reprogramming Demystified: A Neural Tangent Kernel Perspective
Ming-Yu Chung, Jiashuo Fan, Hancheng Ye, Qinsi Wang, Wei Shen, Chia-Mu Yu, Pin-Yu Chen, Sy-Yen Kuo
arXiv:2506.00620 · Preprint
2025
Tissue Reassembly with Generative AI
Tingyang Yu, Chanakya Ekbote, Nikita Morozov, Jiashuo Fan, et al., Pascal Frossard, Maria Brbić
bioRxiv · 10.1101/2025.02.13.638045
2023
RCA-NOC: Relative Contrastive Alignment for Novel Object Captioning
Jiashuo Fan, Yaoyuan Liang, Leyao Liu, Shao-Lun Huang, Lei Zhang
ICCV 2023 — IEEE/CVF International Conference on Computer Vision
2023
LUNA: Language as Continuing Anchors for Referring Expression Comprehension
Yaoyuan Liang, Zhao Yang, Yansong Tang, Jiashuo Fan, Ziran Li, Jingang Wang, Philip H. S. Torr, Shao-Lun Huang
ACM Multimedia 2023
2023
ClickSeg: 3D Instance Segmentation with Click-Level Weak Annotations
Leyao Liu, Tao Kong, Minzhao Zhu, Jiashuo Fan, Lu Fang
arXiv:2307.09732 · Preprint
2022
UCC: Uncertainty Guided Cross-Head Co-Training for Semi-Supervised Semantic Segmentation
Jiashuo Fan, Bin-Bin Gao, Huan Jin, Lihui Jiang
CVPR 2022 · DOI · arXiv