Details

Autor(en) / Beteiligte

• El-Shafai, Walid
• El-Fattah, Ibrahim Abd
• Taha, Taha E.

Titel

Deep learning-based hair removal for improved diagnostics of skin diseases

Ist Teil von

• Multimedia tools and applications, 2024-03, Vol.83 (9), p.27331-27355

Ort / Verlag

New York: Springer US

Erscheinungsjahr

2024

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• The incidence of melanoma, the most serious form of skin cancer, has been increasing rapidly in recent years. Early diagnosis is crucial for successful treatment. Dermoscopy, a reliable medical technique, utilizes specialized devices to examine the skin and detect melanoma. With advancements in digital imaging, high-quality images of these examinations can now be captured and stored. These images are being standardized and used for automated melanoma detection. However, the presence of hair on the skin poses a challenge to accurate diagnosis. Thus, it is essential to remove hair to obtain precise results. In this paper, we propose a simple yet effective method for hair removal using deep learning. Our approach leverages the architecture of generative adversarial networks (GAN) combined with convolutional neural networks (CNN) to reconstruct hair-free images. The GAN consists of a generator and a discriminator. The generator takes a dermoscopy image as input and aims to generate a latent distribution that eliminates hair, considering it as noise. Simultaneously, the discriminator detects changes in the generated image. This iterative process continues until the discriminator fails to identify any changes, considering the generated image as the original hairless image. To evaluate our proposed model, a dataset comprising both hair-covered and hairless images is required. As such a dataset does not currently exist, we introduce a new dataset called Modified-HAM10000 (M-HAM10000), inspired by the scientifically curated dermoscopy dataset HAM10000. Experimental results demonstrate the improved performance of our technique on the M-HAM10000 dataset. Furthermore, we employ various evaluation metrics including Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Multiscale Structural Similarity Index (MS-SSIM) to assess our model's effectiveness. Through experiments conducted on the publicly available M-HAM10000 dataset, our proposed method demonstrates high efficiency in hair removal, enhancing the accuracy of skin disease diagnostics compared to other existing methods.