Enhanced Dermatology Image Quality Assessment via Cross-Domain Training

Authors

Ignacio Hernández Montilla, Alfonso Medela, Paola Pasquali, Andy Aguilar, Taig Mac Carthy, Gerardo Fernández, Antonio Martorell, Enrique Onieva.

Introduction

Image quality is a critical factor in dermatological AI applications. Poor-quality images can lead to incorrect diagnoses, unreliable severity assessments, and failed telemedicine consultations. While our previous work on Dermatology Image Quality Assessment (DIQA) addressed this problem, the challenge of generalizing quality assessment across different imaging conditions, devices, and clinical settings remained.

This research presents an enhanced approach to dermatology image quality assessment that leverages cross-domain training techniques to improve model generalization across diverse clinical environments.

The Challenge of Domain Shift

Dermatological images are captured using a wide variety of devices—from professional medical cameras to consumer smartphones—under vastly different lighting conditions. This creates significant domain shift between training and deployment environments:

• Device variability: Different sensors, lenses, and image processing pipelines
• Lighting conditions: Clinical lighting vs. natural light vs. artificial indoor lighting
• Image capture protocols: Standardized clinical photography vs. patient self-captures
• Skin types: Variations in melanin content affect image characteristics

Traditional deep learning models trained on a single domain often fail to generalize to these varied conditions, limiting their real-world applicability.

Method

Our enhanced DIQA approach addresses domain shift through:

1. Multi-source domain training: Training on images from multiple clinical sites, devices, and capture conditions
2. Domain-invariant feature learning: Extracting quality-relevant features that transfer across domains
3. Adversarial domain adaptation: Reducing the distribution gap between source and target domains
4. Self-supervised pretraining: Learning robust representations from unlabeled dermatological images

Results

The cross-domain training approach demonstrates significant improvements over single-domain baselines:

• Improved generalization to unseen devices and clinical settings
• Reduced performance degradation when deployed in new environments
• Consistent quality assessment across diverse patient populations
• Better calibration of quality scores across domains

These improvements make the enhanced DIQA system more suitable for deployment in real-world clinical and telemedicine applications where image capture conditions cannot be controlled.

Impact

By improving the robustness of dermatological image quality assessment, this work enables:

• More reliable AI-assisted diagnosis in telemedicine
• Consistent image quality standards across multi-site clinical trials
• Better patient experience through real-time feedback on image quality
• Reduced need for repeated image captures and follow-up consultations

Read full text

Available at https://doi.org/10.1145/3774976.3774977.