A Two-Stage End-to-End Deep Learning Framework for Pathologic Examination in Skin Tumor Diagnosis.

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in medicine

Background:

• Pathologic examination of skin tumors is crucial but labor-intensive.
• Digitization and AI offer potential to enhance diagnostic efficiency.
• Neurofibromas (NF), Bowen disease (BD), and seborrheic keratosis (SK) are common skin tumors requiring accurate diagnosis.

Purpose of the Study:

• To develop an AI-driven framework for diagnosing NF, BD, and SK from digitized pathology slides.
• To improve the efficiency and accuracy of skin tumor diagnosis.
• To establish an end-to-end, extendable system for computational pathology.

Main Methods:

• A two-stage framework involving patches-wise and slide-wise diagnosis.
• Convolutional neural networks (CNNs) for feature extraction from image patches.
• Attention graph gated network combined with post-processing for slide-level diagnosis.

Main Results:

• The framework was trained and validated on NF, BD, SK, and negative samples.
• Classification performance was evaluated using accuracy and receiver operating characteristic curves.
• Demonstrated feasibility of deep learning for diagnosing these specific skin tumors.

Conclusions:

• The proposed AI framework shows promise for efficient and accurate skin tumor diagnosis.
• This study is a pioneering application of deep learning to NF, BD, and SK pathology.
• The framework has the potential to streamline the diagnostic workflow in dermatopathology.