Enhancing Skin Lesion Detection: A Multistage Multiclass Convolutional Neural Network-Based Framework.

Area of Science:

• Dermatology
• Computer-aided diagnosis
• Machine learning

Background:

• Accurate skin lesion detection is crucial for timely dermatological treatment.
• Machine learning and computer-aided diagnosis offer promising advancements in analyzing skin lesions.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN)-based multistage framework for classifying seven types of skin lesions.
• To improve the accuracy and efficiency of skin lesion classification using transfer learning.

Main Methods:

• A two-stage CNN model was designed, first classifying lesions as benign or malignant.
• Transfer learning was applied in the second stage to subclassify benign and malignant lesions into specific types.
• Frozen weights from the initial CNN training were utilized for efficient transfer learning.

Main Results:

• The framework achieved 93.4% accuracy in classifying benign and malignant lesions on the ISIC2018 dataset.
• Subclassification accuracy reached 96.2% for both benign and malignant lesion types.
• The proposed method demonstrated superior classification rates and reduced training time compared to existing CNN models.

Conclusions:

• The multistage, multiclass CNN framework effectively categorizes diverse skin lesions.
• This approach enhances diagnostic accuracy and efficiency in dermatological applications.
• The study highlights the potential of transfer learning in deep learning for medical image analysis.