Digital hair segmentation using hybrid convolutional and recurrent neural networks architecture

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Melanoma diagnosis relies on visual inspection, with hair obscuring skin lesions.
• Digital hair removal is a key pre-processing step, involving segmentation and inpainting.
• Accurate hair segmentation is challenging due to varied hair characteristics and lesion textures.

Purpose of the Study:

• To develop a deep learning approach for automated hair segmentation in skin lesion images.
• To improve the accuracy and robustness of hair removal as a pre-processing step for melanoma detection.
• To address limitations of traditional thresholding methods in hair segmentation.

Main Methods:

• A hybrid deep learning network combining convolutional and recurrent layers was proposed.
• Weakly labeled data was utilized for training the segmentation model.
• Encoded features were processed by recurrent layers to capture spatial dependencies.
• Evaluation used the Jaccard Index and a tumour disturb pattern metric on a public dataset.

Main Results:

• The proposed hybrid network achieved a Jaccard Index of 77.8%, outperforming the state-of-the-art method (66.5%).
• The tumour disturb pattern was reduced to 14%, compared to 23% for the existing method.
• Superior segmentation accuracy and reduced impact on lesion texture were demonstrated.

Conclusions:

• The hybrid deep learning architecture effectively segments hair from skin and lesions.
• Weakly labeled ground truth was sufficient for training an accurate hair segmentation model.
• The method offers a robust solution for hair removal in dermatological imaging.