GP-CNN-DTEL: Global-Part CNN Model With Data-Transformed Ensemble Learning for Skin Lesion Classification

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Accurate skin lesion classification is hindered by similar appearances across different lesion types and variations within the same type.
• Single Deep Convolutional Neural Networks (Deep CNNs) often exhibit poor generalization when trained on limited datasets.

Purpose of the Study:

• To develop an advanced model for precise skin lesion classification.
• To address the limitations of existing methods in handling inter-class similarity and intra-class variation.
• To improve the generalization ability of deep learning models for medical image analysis.

Main Methods:

• Proposed a Global-Part Convolutional Neural Network (GP-CNN) model integrating Global CNN (G-CNN) for overall image context and Part CNN (P-CNN) for local details.
• G-CNN utilized downscaled images to generate Classification Activation Maps (CAMs).
• P-CNN employed CAM-guided cropped patches for detailed region analysis, enhanced by a data-transformed ensemble learning strategy.

Main Results:

• The GP-CNN model achieved state-of-the-art performance on the ISIC 2016 and ISIC 2017 Skin Lesion Challenge (SLC) datasets.
• Achieved an Area Under the Precision-Recall Curve (AP) of 0.718 on ISIC 2016 and an Average Area Under the ROC Curve (AUC) of 0.926 on ISIC 2017.
• Demonstrated superior classification accuracy without relying on external datasets.

Conclusions:

• The GP-CNN model effectively balances global and local feature extraction for improved skin lesion classification.
• The ensemble learning strategy further boosts performance by integrating diverse data transformations.
• The proposed method offers a robust and data-efficient solution for automated skin lesion diagnosis.