Abstract
Transfer learning addresses limited labeled data in deep learning by leveraging pre-trained model knowledge. In extreme data scarcity scenarios, such as rare medical conditions, few-shot classification becomes essential, relying critically on feature representation quality. A key challenge is producing discriminative, transferable representations when target domains diverge significantly from source domains. This study investigates strategies to improve pre-trained model representations for few-shot classification of skin-related Neglected Tropical Diseases (NTDs) on dark skin tones (Fitzpatrick Scale IV–VI). We propose a sequential framework combining supervised domain adaptation to align representations with darker skin dermatology images, followed by supervised contrastive learning (SupCon) to enhance inter-class separability under data scarcity. Using DenseNet-161 and ViT-B/16 as feature extractors, we evaluated performance over 100 randomly sampled support/query episodes per setting and reported mean accuracy with 95% confidence intervals and paired significance tests. Domain adaptation significantly improved the pre-trained baseline for ViT-B/16 across all five few-shot settings (Wilcoxon, p < 0.001) and for DenseNet-161 in three of five settings (p < 0.05). The highest mean accuracy, 81.1% (95% CI [78.9, 83.3]), was achieved by ViT-B/16 in the 3-way 5-shot setting after supervised contrastive learning. Although its additional gains beyond domain adaptation were generally modest, supervised contrastive learning complemented the pipeline by refining learned feature representations and achieving the highest mean performance in a select few-shot settings. These findings demonstrate the effectiveness of combining domain adaptation with supervised contrastive learning to improve few-shot NTD recognition on dark skin tones.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=sx0JslEg32
BibTex
@InProceedings{ObeFel_Improving_MICCAISAT2026,
author = { Obete, Felix AND Osolo, Ian Raymond AND Michael-Ahile, Terhemba},
title = { { Improving Feature Representations for Few-shot Classification of Neglected Tropical Diseases on Dark Skin Tones } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}