Enhanced Skin Lesion Classification through Hair Artifact Removal and Transfer Learning Models

Kerem Gencer

Makale detayları

Enhanced Skin Lesion Classification through Hair Artifact Removal and Transfer Learning Models

Yazarlar:

Kerem Gencer

Yayıncı:

Karyay Karadeniz Yayımcılık

Ana Başlıklar:

Biyomedikal Görüntüleme

Dergi:

Black Sea Journal of Engineering and Science Vol.8 Issue 4, pp.1007-1021

Belge Tipi:

Makale

Yayın Tarihi:

7/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001346260

pISSN:

2619-8991

Anahtar kelimeler:

Özet:

 Skin cancer is one of the most common types of cancer worldwide and early detection is vital for the most effective treatment. However, melanomas and other malignant skin lesions have been very difficult to detect and this is often because dermatologists use dermoscopy images. These images often contained noise, such as hair and air bubbles, which made diagnosis seriously difficult. To address this important issue, we developed an effective noise removal algorithm to improve the quality of these images using conventional image processing techniques. Using the ISIC2018 dataset, a standard for skin lesion classification, we integrated our de-noising algorithm to carefully pre-process the dermoscopy images. Our method involved selecting the red channel, carefully applying the top hat transform, carefully using Gaussian blur and applying the Otsu thresholding method to detect and remove hairs, followed by internal inpainting to fill the resulting gaps. Following preprocessing, we carefully applied and compared several transfer learning models to classify skin lesions: AlexNet, VGG16, ResNet50 and EfficientNetB0. Our results showed significant improvement in classification accuracy: AlexNet achieved 95.32, VGG16 97.55, ResNet50 99.04 and EfficientNetB0 99.89 accuracy. These findings demonstrated that combining hair removal preprocessing with transfer learning significantly improves the accuracy of automatic skin lesion classification. This innovative approach has the potential to greatly assist dermatologists in detecting skin cancer early and lead to better patient outcomes. Our method is robust and effective and is a valuable tool for improving dermatologic assessments in clinical settings.

 Skin cancer is one of the most common types of cancer worldwide and early detection is vital for the most effective treatment. However, melanomas and other malignant skin lesions have been very difficult to detect and this is often because dermatologists use dermoscopy images. These images often contained noise, such as hair and air bubbles, which made diagnosis seriously difficult. To address this important issue, we developed an effective noise removal algorithm to improve the quality of these images using conventional image processing techniques. Using the ISIC2018 dataset, a standard for skin lesion classification, we integrated our de-noising algorithm to carefully pre-process the dermoscopy images. Our method involved selecting the red channel, carefully applying the top hat transform, carefully using Gaussian blur and applying the Otsu thresholding method to detect and remove hairs, followed by internal inpainting to fill the resulting gaps. Following preprocessing, we carefully applied and compared several transfer learning models to classify skin lesions: AlexNet, VGG16, ResNet50 and EfficientNetB0. Our results showed significant improvement in classification accuracy: AlexNet achieved 95.32, VGG16 97.55, ResNet50 99.04 and EfficientNetB0 99.89 accuracy. These findings demonstrated that combining hair removal preprocessing with transfer learning significantly improves the accuracy of automatic skin lesion classification. This innovative approach has the potential to greatly assist dermatologists in detecting skin cancer early and lead to better patient outcomes. Our method is robust and effective and is a valuable tool for improving dermatologic assessments in clinical settings.

 Skin cancer is one of the most common types of cancer worldwide and early detection is vital for the most effective treatment. However, melanomas and other malignant skin lesions have been very difficult to detect and this is often because dermatologists use dermoscopy images. These images often contained noise, such as hair and air bubbles, which made diagnosis seriously difficult. To address this important issue, we developed an effective noise removal algorithm to improve the quality of these images using conventional image processing techniques. Using the ISIC2018 dataset, a standard for skin lesion classification, we integrated our de-noising algorithm to carefully pre-process the dermoscopy images. Our method involved selecting the red channel, carefully applying the top hat transform, carefully using Gaussian blur and applying the Otsu thresholding method to detect and remove hairs, followed by internal inpainting to fill the resulting gaps. Following preprocessing, we carefully applied and compared several transfer learning models to classify skin lesions: AlexNet, VGG16, ResNet50 and EfficientNetB0. Our results showed significant improvement in classification accuracy: AlexNet achieved 95.32, VGG16 97.55, ResNet50 99.04 and EfficientNetB0 99.89 accuracy. These findings demonstrated that combining hair removal preprocessing with transfer learning significantly improves the accuracy of automatic skin lesion classification. This innovative approach has the potential to greatly assist dermatologists in detecting skin cancer early and lead to better patient outcomes. Our method is robust and effective and is a valuable tool for improving dermatologic assessments in clinical settings.

 Skin cancer is one of the most common types of cancer worldwide and early detection is vital for the most effective treatment. However, melanomas and other malignant skin lesions have been very difficult to detect and this is often because dermatologists use dermoscopy images. These images often contained noise, such as hair and air bubbles, which made diagnosis seriously difficult. To address this important issue, we developed an effective noise removal algorithm to improve the quality of these images using conventional image processing techniques. Using the ISIC2018 dataset, a standard for skin lesion classification, we integrated our de-noising algorithm to carefully pre-process the dermoscopy images. Our method involved selecting the red channel, carefully applying the top hat transform, carefully using Gaussian blur and applying the Otsu thresholding method to detect and remove hairs, followed by internal inpainting to fill the resulting gaps. Following preprocessing, we carefully applied and compared several transfer learning models to classify skin lesions: AlexNet, VGG16, ResNet50 and EfficientNetB0. Our results showed significant improvement in classification accuracy: AlexNet achieved 95.32, VGG16 97.55, ResNet50 99.04 and EfficientNetB0 99.89 accuracy. These findings demonstrated that combining hair removal preprocessing with transfer learning significantly improves the accuracy of automatic skin lesion classification. This innovative approach has the potential to greatly assist dermatologists in detecting skin cancer early and lead to better patient outcomes. Our method is robust and effective and is a valuable tool for improving dermatologic assessments in clinical settings.

Makale detayları

Enhanced Skin Lesion Classification through Hair Artifact Removal and Transfer Learning Models

Yazarlar:

Kerem Gencer

Yayıncı:

Karyay Karadeniz Yayımcılık

Ana Başlıklar:

Biyomedikal Görüntüleme

Dergi:

Black Sea Journal of Engineering and Science Vol.8 Issue 4, pp.1007-1021

Belge Tipi:

Makale

Yayın Tarihi:

7/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001346260

pISSN:

2619-8991

Skin cancer

Dermoscopy images

Noise removal algorithm

Transfer learning

Classification

Image processing techniquesSkin cancer

Dermoscopy images

Noise removal algorithm

Transfer learning

Classification

Image processing techniques

Sayı

Makale detayları

Yazarlar: Kerem Gencer

Yayıncı: Karyay Karadeniz Yayımcılık

Ana Başlıklar: Biyomedikal Görüntüleme

Dergi: Black Sea Journal of Engineering and Science Vol.8 Issue 4, pp.1007-1021

Belge Tipi: Makale

Yayın Tarihi: 7/2025

Belge Dili: İngilizce

Document ID (Org No.): 2001346260

pISSN: 2619-8991

Sayı

Yazarlar:

Kerem Gencer

Yayıncı:

Karyay Karadeniz Yayımcılık

Ana Başlıklar:

Biyomedikal Görüntüleme

Dergi:

Black Sea Journal of Engineering and Science Vol.8 Issue 4, pp.1007-1021

Belge Tipi:

Makale

Yayın Tarihi:

7/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001346260

pISSN:

2619-8991