Deep learning-level melanoma detection by interpretable machine learning and imaging biomarker cues.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostics

Background:

• Early melanoma diagnosis is challenging for physicians due to difficulties distinguishing benign from malignant lesions.
• Current deep learning diagnostic tools lack transparency, hindering widespread clinical adoption.
• There is a need for interpretable machine learning in medical image analysis.

Purpose of the Study:

• To develop a transparent machine learning technology for differentiating melanomas from nevi in dermoscopy images.
• To create a user interface for integrating sensory cues into the diagnostic process.
• To compare the performance of interpretable machine learning against deep learning for melanoma detection.

Main Methods:

• An ensemble machine learning classifier (Eclass) was trained using imaging biomarker cues (IBCs).
• A deep learning classifier was trained using raw dermoscopy images.
• Performance was evaluated by comparing areas under the diagnostic receiver operator curves.

Main Results:

• The interpretable machine learning algorithm (Eclass) outperformed the leading deep-learning approach in 75% of comparisons.
• The developed user interface exclusively displayed diagnostic imaging biomarkers as IBCs.
• Eclass demonstrated superior diagnostic accuracy compared to black-box deep learning models.

Conclusions:

• The interpretable Eclass model offers a more physician-friendly alternative to deep learning for melanoma diagnosis.
• Imaging biomarker cues can be integrated into clinical screening workflows.
• This approach has potential applications in other image-based diagnostic fields like pathology and radiology.