Radiomic and deep learning analysis of dermoscopic images for skin lesion pattern decoding

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Noninvasive diagnosis of skin lesions is crucial for early detection of conditions like melanoma.
• Traditional diagnostic methods can be limited by subjectivity and accessibility.
• Advancements in artificial intelligence and image analysis offer potential for improved diagnostic accuracy.

Purpose of the Study:

• To evaluate the efficacy of a hybrid deep learning and radiomics model for diagnosing skin lesions using dermoscopic images.
• To integrate patient metadata with imaging analysis for enhanced diagnostic performance.
• To compare the hybrid model's performance against existing classification methods.

Main Methods:

• Utilized the International Skin Imaging Collaboration (ISIC) dataset (2016-2020) containing diverse skin lesions.
• Developed a hybrid model combining deep learning with comprehensive radiomics feature extraction.
• Quantified skin lesion patterns using a wide array of image features and patient metadata.
• Benchmarked the model against seven ISIC 2020 challenge methods using a binary and multiclass framework.

Main Results:

• The hybrid model achieved superior performance in distinguishing benign from malignant lesions.
• Achieved high Area Under the Receiver Operating Characteristic Curve (AUROC) scores: 99% (internal ISIC 2018), 95% (Jinan dataset), and 96% (Longhua dataset).
• Demonstrated high sensitivity (97.6%, 93.9%, 96.0%) and specificity (98.4%, 96.7%, 96.9%) across datasets.

Conclusions:

• The integration of radiomics and deep learning effectively captures skin lesion heterogeneity and pattern expression.
• This hybrid approach offers a promising noninvasive tool for accurate skin lesion diagnosis.
• The findings support the potential of AI-driven analysis of dermoscopic images for improved dermatological care.