UC Santa Cruz

Medical image analysis is constrained by a structural mismatch: clinical decision-making demands fine-grained understanding across diverse populations, body sites, and disease severities, yet expert annotations are scarce, demographic coverage is uneven, and rare presentations are nearly absent. This dissertation develops deep learning algorithms that compensate for that mismatch through four connected works in dermatology and wound care, domains where the limitations are most acute, and the downstream stakes for under-imaged populations are highest. We develop four methods spanning unsupervised, self-supervised, and supervised strategies, with both non-parametric and parametric generative paradigms. When there are no labels, we learn unsupervised from temporal structure (HealNet); when there is no temporal structure, we learn supervised from a frozen diffusion backbone (FEDD); when the data themselves are too few and too imbalanced, we generate them with sensitive-attribute control (cgDDI); and to close the loop into 3D space, we develop metric scale dermatology foundation models with non-parametric synthetic and real data (DermDepth).HealNet introduces a self-supervised framework for acute wound heal-stage classification, learning biologically meaningful temporal embeddings without any human-provided labels and reaching 97.7% pre-text and 90.6% downstream stage-classification accuracy on a small longitudinal cohort. FEDD generalizes representation learning to the single-image clinical setting by re-tasking a frozen denoising-diffusion backbone as the encoder, achieving state-of-the-art segmentation (IoU lift 0.06–0.18) and a 14-point gain in malignancy classification accuracy on the biopsy-confirmed Diverse Dermatology Images (DDI) benchmark while using only 5–20% of the available labels, and contributes the first explicit skin-tone-stratified fairness evaluation of the work. cgDDI repurposes that diffusion backbone as a controllable generator, combining latent-diffusion in-painting, non-parametric lesion mapping, and textual-inversion plus LoRA prior-preservation to grow a 656-image dataset by over 400×, balancing the full Fitzpatrick spectrum and lifting DDI malignancy classification to state-of-the-art 90.9% accuracy with EOM fairness 86.6 (up from 69.6) and a +13.9% cross-dataset gain on Fitzpatrick17k. DermDepth closes the loop from 2D to 3D, presenting the first single-view metric-scale 3D reconstruction model for dermatology together with D-Synth, the first synthetic dermoscopic dataset with pixel-perfect depth, normals, and camera intrinsic; a 2.1 M-parameter scale-and-normal head on top of a frozen MoGe-2 backbone corrects metric scale error from over 16× to under 1.1× on real dermoscopic data, reduces skin-tone scale disparity from 10.90 to 1.02, and validates 3D lesion width, area, and volume estimation across SKINL2, WoundsDB, and DDI against the published clinical literature.Across the work, a pattern recurs: small, clean datasets; frozen large backbones with trainable heads; and explicit, skin-tone-stratified fairness evaluation. We believe this framework generalizes beyond dermatology to any data-constrained medical imaging domain. All synthetic data, fine-tuned checkpoints, annotation work, code, and per-disease generative models are released openly to support reproducibility and continued fairness research.