[Paper] Uncertainty-Aware Domain Adaptation for Vitiligo Segmentation in Clinical Photographs

Source: arXiv - 2512.11791v1

Overview

The paper introduces a trustworthy, uncertainty‑aware segmentation pipeline for automatically measuring vitiligo lesions in routine clinical photographs. By blending domain‑adaptive pre‑training, a high‑frequency‑focused ConvNeXt V2 backbone, and ensemble‑based uncertainty estimation, the authors achieve state‑of‑the‑art accuracy while giving clinicians clear signals about where the model is unsure.

Key Contributions

• Domain‑adaptive pre‑training on the large ISIC 2019 skin‑lesion dataset, followed by a region‑of‑interest (ROI) constrained dual‑task loss that suppresses background noise in vitiligo images.
• High‑Frequency Spectral Gating (HFSG) module and stem‑skip connections integrated into a ConvNeXt V2 encoder, enabling the network to capture the subtle texture differences that characterize vitiligo borders.
• Uncertainty quantification via a K‑fold ensemble combined with Test‑Time Augmentation (TTA), producing pixel‑wise entropy maps that flag ambiguous regions for clinician review.
• Comprehensive validation on an expert‑annotated clinical cohort, reporting a Dice score of 85.05 % and a 30 % reduction in 95 % Hausdorff Distance compared to strong CNN and Transformer baselines.
• Zero catastrophic failures and interpretable entropy visualizations, establishing a practical reliability standard for automated dermatology tools.

Methodology

1. Data‑efficient training

   • Start with a domain‑adaptive pre‑training phase on the ISIC 2019 dataset (large, publicly available skin‑lesion images).
   • Fine‑tune on the vitiligo cohort using a dual‑task loss: (i) a standard segmentation loss (e.g., Dice/CE) and (ii) an ROI‑constrained auxiliary loss that penalizes predictions outside the clinician‑marked region of interest, effectively filtering out background clutter.

2. Network architecture

   • The backbone is ConvNeXt V2, a modern convolutional architecture that balances speed and expressiveness.
   • A High‑Frequency Spectral Gating (HFSG) block processes feature maps in the frequency domain, attenuating low‑frequency components while preserving high‑frequency edges that delineate vitiligo patches.
   • Stem‑skip connections shuttle early‑stage, high‑resolution features directly to deeper decoder stages, helping the model retain fine‑grained texture cues.

3. Uncertainty & trust layer

   • Train K (e.g., 5) independent models on different data folds.
   • At inference, apply Test‑Time Augmentation (rotations, flips, color jitter) to each model, aggregate the softmax outputs, and compute pixel‑wise entropy as an uncertainty score.
   • High entropy regions are highlighted for manual review, turning the black‑box model into an assistive tool rather than a fully autonomous decision maker.

Results & Findings

• The HFSG module contributed ~3 % Dice gain and a 5 px reduction in Hausdorff distance.
• Ensemble + TTA lowered average entropy in clear‑cut regions while preserving high entropy on ambiguous borders, aligning with dermatologist feedback.

Practical Implications

• Clinical workflow integration – The uncertainty maps can be overlaid on existing electronic health record (EHR) viewers, letting dermatologists focus their attention on the most ambiguous lesion edges, thus speeding up charting and longitudinal tracking.
• Scalable deployment – The ConvNeXt V2 + HFSG backbone runs comfortably on a single mid‑range GPU (e.g., RTX 3060), making it feasible for on‑premise hospital servers or edge devices in tele‑dermatology kits.
• Regulatory friendliness – Zero catastrophic failures and explicit uncertainty reporting satisfy emerging AI‑medical device guidelines that demand “human‑in‑the‑loop” safety nets.
• Transferability – The domain‑adaptive pre‑training + ROI‑constrained loss recipe can be repurposed for other skin‑condition segmentation tasks (e.g., psoriasis, melasma) where background clutter is a major challenge.

Limitations & Future Work

• Dataset diversity – The clinical cohort, while expert‑annotated, comprises images from a limited number of clinics and lighting conditions; broader multi‑center validation is needed.
• Computation overhead – K‑fold ensembles and TTA increase inference latency; future work could explore Bayesian approximations or lightweight uncertainty heads to retain speed.
• Fine‑grained lesion classification – The current pipeline segments vitiligo but does not differentiate active vs. stable lesions; integrating a classification head could support treatment decision‑making.

Bottom line: By marrying frequency‑aware feature extraction with rigorous uncertainty estimation, this work delivers a practical, trustworthy tool for automated vitiligo assessment—an approach that can inspire similar solutions across the broader medical imaging landscape.

Authors

• Wentao Jiang
• Vamsi Varra
• Caitlin Perez‑Stable
• Harrison Zhu
• Meredith Apicella
• Nicole Nyamongo

Paper Information

• arXiv ID: 2512.11791v1
• Categories: cs.CV
• Published: December 12, 2025
• PDF: Download PDF