Carrión, H. (CSE) – Deep Learning Algorithms for Medical Image Representation Learning and Understanding

AI-assisted clinical decisions in medicine, and particularly in dermatology, demand fine-grained understanding across diverse skin tones, body sites, and disease types, yet expert-annotated datasets are scarce, demographically imbalanced, and almost devoid of rare presentations. This dissertation develops four deep learning systems for this low-label, low-coverage regime. We introduce HealNet, which learns wound healing stages from longitudinal photographs without any human labels, reaching 90.6% downstream stage-classification accuracy on a small longitudinal cohort. The Fair, Efficient, and Diverse Diffusion (FEDD) model then leverages powerful diffusion-model embeddings to build a skin-tone-fair, data-efficient classifier for skin lesions, matching or exceeding state-of-the-art performance while using only 5-20% of available labels and contributing explicit skin-tone-stratified fairness evaluation of the work. Next, Controllable Generation of Diverse Dermatological Imagery (cgDDI) re-tasks this diffusion model to controllably synthesize skin-tone-balanced dermatological imagery, growing a small biopsy-confirmed dataset by over 400x and reaching state-of-the-art 90.9% accuracy and improved fairness in malignancy classification, with a +13.9% cross-dataset gain on the Fitzpatrick17k benchmark. Finally, we introduce D-Synth and DermDepth: a synthetic dermoscopic dataset with pixel-perfect 3D ground truth and a metric-scale foundation model that closes the loop into 3D dermatology, correcting metric scale error from over 16x to under 1.1x on real dermoscopic data and enabling single-photograph measurement of lesion reconstruction: size, area, and volume without specialized hardware. All data, code, and models are released openly to support reproducibility and ongoing fairness research.

Event Host:  Héctor Carrión, Ph.D. Candidate, Computer Science & Engineering

Advisor: Narges Norouzi

Zoom: https://ucsc.zoom.us/j/96678782408?pwd=71f0ObEnUMNgkZ9NYnpbFLMlg1Pdm0.1

Passcode: 0FMVtz