Analyzing the Impact of Image Denoising and Segmentation on Melanoma Classification Using Convolutional Neural...

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Early detection and treatment of skin cancer are vital for global mortality reduction.
• Deep learning models, particularly convolutional neural networks (CNNs), show promise for automated skin lesion detection and classification.

Purpose of the Study:

• To investigate the impact of pre-processing techniques (data augmentation, contrast enhancement, segmentation) on CNN performance for skin lesion classification.
• To design and evaluate a novel CNN architecture for improved lesion classification.
• To compare the proposed network's performance against state-of-the-art methods on a benchmark dataset.

Main Methods:

• A custom CNN architecture was developed with unique layer organization, kernel counts, network depth, and hyperparameters.
• Pre-processing steps including denoising, data augmentation, contrast enhancement, and segmentation were applied to the HAM10000 dataset.
• The CNN model was trained and evaluated using the pre-processed and segmented data.

Main Results:

• The proposed CNN, utilizing denoised and segmented data, achieved high performance metrics.
• Specific performance scores included: Accuracy (ACC) 93.40%, Precision (PRE) 93.45%, Recall (REC) 94.51%, Specificity (SPE) 92.08%, and F-score 93.98%.

Conclusions:

• Pre-processing and segmentation steps significantly improve CNN performance for skin lesion classification.
• The developed CNN demonstrates competitive performance compared to existing state-of-the-art methods.
• Incorporating these data enhancement techniques is crucial for advancing automated skin cancer detection systems.