Skip to content
Kean CS Lab

Program 02 / 06

Vision, Imaging & Health AI

Representation learning for images — from medical volumes to faces that last a fraction of a second.

Computer vision at Kean spans medical image analysis, general deep learning for vision, and affective computing. The shared discipline is evaluation: a segmentation that looks clean on a challenge set is not a clinical claim, and a microexpression classifier is not a theory of emotion until the dataset and protocol say so.

Questions the work is for

  1. 01

    Which inductive biases survive domain shift between scanners, sites, and patient populations?

  2. 02

    How do we learn from scarce, privacy-constrained medical labels?

  3. 03

    What does a fair, reproducible protocol look like for fleeting facial behavior?

Methods, data, benchmarks

method

Medical image analysis

Convolutional and transformer backbones for detection, segmentation, and quantification; careful train/val/test hygiene across sites.

dataset

Microexpression corpora

Public review work from Kwak and collaborators surveys recognition, classification, and datasets — the map a new doctoral project should start from, not skip.

benchmark

Challenge benchmarks

Standard vision suites (classification, detection, segmentation) plus domain-specific medical challenges. We treat leaderboard position as a starting measurement, not a result.

tooling

On-device vision

Embedded and edge inference (Li, Lyu) when the camera cannot assume a datacenter — latency, memory, and privacy as first-class constraints.

Taxonomy on this page: method, dataset, benchmark, tooling. Field-standard methods are named as such; they are not automatic claims of a Kean dataset release.

Doctoral tracks

Medical imaging

Xu’s public focus is computer vision, deep learning, and medical image analysis. Doctoral work here should name the clinical decision, the imaging modality, and the failure mode — not only the architecture.

Affective and fine-grained vision

Kwak’s microexpression survey is a methods paper in the best sense: it forces a student to choose a dataset, a split, and a claim that the literature has not already exhausted.

Vision + language on the edge

Boyang Li and Kuan Huang work across vision, NLP, and embedded ML. A dissertation can compress multimodal models without pretending compression is free.

Who this is for

Ideal if you already write PyTorch (or equivalent), can read a medical or vision paper to the loss function, and want a project with a measurable artifact — a model, a protocol, or a dataset critique.

Ph.D. Program →

Faculty in this program

Selected papers

← All programs