TATL: Task agnostic transfer learning for skin attributes detection.

Area of Science:

• Dermatology
• Medical Imaging Analysis
• Computer Vision
• Machine Learning

Background:

• Current skin attribute detection methods often use ImageNet pre-trained networks, which are suboptimal for distinct medical datasets with limited samples.
• This limitation hinders the performance of deep learning models in specialized dermatological applications.

Purpose of the Study:

• To propose a novel transfer learning framework, Task Agnostic Transfer Learning (TATL), for improved skin attribute detection.
• To address the challenges of limited training data and domain shift in medical image analysis.

Main Methods:

• Developed TATL, a framework that first learns an attribute-agnostic segmenter for general lesion skin region detection.
• Transferred knowledge from the segmenter to attribute-specific classifiers for detecting individual skin attributes.
• Evaluated TATL using various neural network architectures on two benchmark datasets for skin attribute detection.

Main Results:

• TATL demonstrated robust performance across multiple neural network architectures.
• Achieved state-of-the-art results in skin attribute detection tasks.
• Showcased effectiveness in compensating for limited training data, particularly for rare attributes.

Conclusions:

• TATL offers an effective transfer learning mechanism for skin attribute detection, outperforming traditional fine-tuning approaches.
• The framework provides minimal model and computational complexity while achieving high performance.
• Theoretical insights support the practical efficacy of TATL in dermatological image analysis.