Few-shot skin lesion classification with Adaptive Multi-Scale Convolutional Attention Network
Clinical Snapshot
PICO Framework
| P — Population | Digital dermoscopic and clinical skin lesion images from publicly available datasets (HAM10000 and PAD-UFES-20), representing diverse skin lesion types including melanoma, basal cell carcinoma, and other dermatological conditions |
| I — Intervention | Adaptive Multi-Scale Convolutional Attention Network (AMCANet) — a novel deep learning architecture incorporating adaptive multi-scale convolution, hierarchical channel attention, and skin spatial attention modules, evaluated under few-shot learning conditions |
| C — Comparator | Existing baseline deep learning models for skin lesion classification (specific comparator architectures described within the paper, evaluated on the same datasets under equivalent few-shot conditions) |
| O — Outcomes | Classification accuracy, and other standard multi-class performance metrics (precision, recall, F1-score, AUC) across multiple few-shot experimental configurations on HAM10000 and PAD-UFES-20 datasets |
Bottom Line
This paper presents AMCANet, a novel deep learning architecture for few-shot skin lesion classification, evaluated on two public benchmark datasets (HAM10000 and PAD-UFES-20). The conceptual design — combining adaptive multi-scale convolution with hierarchical attention mechanisms — is well-motivated for the genuine challenges of dermoscopic image analysis, particularly in data-scarce settings. However, the study has critical methodological limitations that preclude clinical translation at this stage. No quantitative results are reported in the abstract, confidence intervals are absent, and clinically essential metrics such as melanoma sensitivity and specificity are not described. The comparator models are unspecified, and no external validation on independent or geographically diverse datasets has been performed. For Australian clinicians, the absence of validation on diverse skin phototypes and Australian clinical imaging contexts is a significant gap given our unique epidemiological burden. This work represents an early-stage computational proof-of-concept that may inform future research directions in AI-assisted dermatology. Senior clinicians should not interpret these findings as evidence supporting clinical deployment. Independent prospective validation studies, ideally incorporating Australian patient cohorts and conducted in partnership with dermatologists, are required before any clinical recommendation can be made.
Key Findings
P Value: Not reported
Effect Size: Not reported in abstract — described qualitatively as 'significantly outperforms existing baseline models across multiple metrics'
Primary Outcome: Multi-class skin lesion classification accuracy and related performance metrics under few-shot learning conditions on HAM10000 and PAD-UFES-20 datasets
Nnt Or Sensitivity: Sensitivity, specificity, and AUC for individual lesion classes (including melanoma) not reported in abstract — critical omission for clinical diagnostic accuracy assessment
Confidence Interval: Not reported
Clinical Application
Clinical feasibility is entirely unassessed. The study does not address integration with existing dermatology information systems, real-time inference requirements, regulatory approval pathways, clinician interface design, or workflow impact. Few-shot learning models require careful validation before deployment to ensure they do not underperform on lesion types underrepresented in training episodes. Australia has the highest age-standardised melanoma incidence globally (AIHW 2023), making robust AI-assisted skin lesion classification highly relevant. However, this model has not been validated on Australian patient data, Aboriginal and Torres Strait Islander skin types, or images acquired in Australian clinical settings (general practice, teledermatology). The TGA regulates AI-based medical devices under the Software as a Medical Device (SaMD) framework, requiring clinical evidence of safety and performance that this study does not provide. RACGP guidelines emphasise dermoscopy training and specialist referral pathways; any AI tool would need to demonstrate added value within these pathways. PBS-listed dermatology consultations and the MBS telehealth framework for remote skin lesion assessment represent potential implementation contexts, but require prospective Australian validation studies first. Potentially applicable as a research prototype for computer-aided detection support in dermoscopy-based skin lesion triage, particularly in settings with limited annotated training data. Not validated for direct clinical use in any population.
Abstract
Computer-aided diagnosis of skin lesions faces core challenges, including scale diversity, blurred boundaries, intra-class morphological variations, and sparse data. Existing methods often rely on fixed receptive fields or generic attention mechanisms, struggling to fully adapt to the unique characteristics of skin lesions. To address this, we propose an Adaptive Multi-scale Convolutional Attention Network (AMCANet), which aims to achieve accurate and robust classification of skin lesions with limited data. AMCANet comprises three core modules: the adaptive multi-scale convolution module dynamically adjusts the receptive field to accommodate lesions of varying sizes; the hierarchical channel attention module integrates multi-level semantic information across different resolutions; and the skin spatial attention module leverages image gradient information to enhance lesion boundaries and local texture features. Extensive few-shot experiments on the HAM10000 and PAD-UFES-20 public datasets demonstrate that AMCANet significantly outperforms existing baseline models across multiple metrics, exhibiting promising generalization capabilities on the evaluated datasets. Qualitative and visual analyses further validate the model's ability to extract discriminative features and effectively focus on lesion regions. This study proposes a deep learning model, which demonstrates certain effectiveness in classifying skin lesions even with a small number of samples, providing a potential direction for future research.
References
- 1.Jin, H., Liu, E., Xu, Q., Li, Y., Chen, X., & Chen, W. (2026). Few-shot skin lesion classification with Adaptive Multi-Scale Convolutional Attention Network. PLOS ONE. https://doi.org/10.1371/journal.pone.0351318
Related Research
Journal of tissue viability
Clinical decision support tools in wound management: A scoping review of existing and emerging tools.
2 Aug 2026
Lasers in medical science
Facial contour modulation and skin tightening using 40.68-MHz unipolar radiofrequency
27 July 2026
Clinical and experimental dermatology
Pemphigus vegetans: insights from a systematic review of the literature.
25 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service