Hair removal in dermoscopy images using variational autoencoders

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Melanoma incidence is increasing, highlighting the need for accurate early detection methods.
• Dermoscopy is crucial for melanoma diagnosis, but hair occlusion in images poses a significant challenge.
• Digital dermoscopy images are increasingly used for automated melanoma detection, necessitating hair removal techniques.

Purpose of the Study:

• To develop an efficient and simple method for hair removal from dermoscopy images.
• To improve the accuracy of automated melanoma detection by eliminating hair artifacts.
• To enhance the visual quality of reconstructed hair-free dermoscopy images.

Main Methods:

• A variational autoencoder (VAE) model was employed for hair removal without requiring paired samples.
• The VAE encoder learns a latent representation that ignores hair, while the decoder reconstructs hair-free images.
• A two-stage training process and three loss functions (SSIM, L1, L2) were utilized to optimize image quality and texture preservation.

Main Results:

• The proposed VAE method successfully generated hair-free dermoscopy images.
• Evaluation using t-distributed stochastic neighbor embedding (SNE) demonstrated the efficiency of the method.
• Experiments on the HAM10000 dataset confirmed the effectiveness of the hair removal technique.

Conclusions:

• The developed variational autoencoder offers a promising solution for hair removal in dermoscopy images.
• This technique can significantly aid in improving the accuracy of computer-aided diagnosis for melanoma.
• The method's ability to preserve image quality makes it suitable for clinical applications.