Skin lesion classification of dermoscopic images using machine learning and convolutional neural network

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Early detection of skin cancer, particularly melanoma, relies on identifying pigmented skin lesions.
• Image detection and computer classification can significantly improve diagnostic accuracy for skin malignancies.

Purpose of the Study:

• To evaluate the effectiveness of data augmentation and k-fold cross-validation in improving skin lesion classification accuracy.
• To compare the performance of Convolutional Neural Networks (CNNs) against traditional Machine Learning algorithms for skin disease detection.

Main Methods:

• Utilized a subset of the HAM10000 dataset (10,015 images) for training and validation.
• Applied data augmentation techniques to enhance model learning of distinguishing features.
• Implemented k-fold cross-validation to ensure model robustness against testing data.
• Analyzed classification accuracy using various Machine Learning algorithms and CNN models.

Main Results:

• Models incorporating data augmentation demonstrated improved learning of distinguishing characteristics.
• K-fold cross-validation enhanced model robustness.
• Convolutional Neural Networks (CNNs) achieved superior classification accuracy compared to other implemented Machine Learning algorithms.
• The highest accuracy achieved was 95.18% using the CNN model.

Conclusions:

• CNNs offer higher accuracy for skin lesion classification than traditional ML algorithms.
• The proposed system, leveraging data augmentation and k-fold cross-validation, facilitates early identification of seven skin disease classes.
• Accurate early diagnosis supports timely medical intervention and treatment by practitioners.