Skin Lesion Analysis By Multi-Target Deep Neural Networks.

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Automatic skin lesion analysis is crucial for early diagnosis.
• Current methods often involve separate steps for segmentation and classification.
• Deep learning models like U-Net and GoogleNet have shown promise but are typically single-task.

Purpose of the Study:

• To develop a single, end-to-end deep neural network model for simultaneous skin lesion segmentation and classification.
• To investigate the effectiveness of a multi-target deep convolutional neural network (DCNN) in analyzing dermoscopic images.
• To explore the commonalities and differences in learning across multiple targets within a single model.

Main Methods:

• A novel multi-target deep convolutional neural network (DCNN) was designed, integrating U-Net and GoogleNet architectures.
• The model was trained to perform three tasks simultaneously: lesion segmentation, melanoma detection, and seborrheic keratosis identification.
• Experiments were conducted using dermoscopic images from the International Skin Imaging Collaboration (ISIC) 2017 Challenge.

Main Results:

• The proposed multi-target DCNN demonstrated superior performance compared to single-target models (U-Net, GoogleNet).
• The model showed significant learning efficiency by handling segmentation and classification tasks concurrently.
• Results indicate the potential of this integrated approach for automated skin lesion diagnosis.

Conclusions:

• A single end-to-end deep neural network model can effectively perform both segmentation and classification of skin lesions simultaneously.
• Multi-target learning in DCNNs offers advantages in learning efficiency and diagnostic potential for skin lesion analysis.
• This work represents a pioneering effort in developing a unified deep learning framework for comprehensive skin lesion assessment.