rcnn

rcnn classifies skin tumors from unprocessed clinical photographs using convolutional neural networks to distinguish benign and malignant lesions for diagnostic assessment.

Key Features:

• Dataset: 40,331 clinical images from 10,426 patients covering 43 distinct disorders collected from January 1, 2008 to March 31, 2019.
• Input data: Uses unprocessed clinical photographs as the imaging input for model inference.
• Algorithmic approach: Employs deep learning with convolutional neural networks (RCNN) for image-based classification tasks.
• Binary classification performance: Achieved an Area Under the Curve (AUC) of 0.863 (95% CI 0.852–0.875) for malignancy detection on unprocessed clinical photographs.
• Sensitivity and specificity (high-sensitivity threshold): Sensitivity 79.1% (76.9%–81.4%) and specificity 76.9% (76.1%–77.8%).
• Sensitivity and specificity (high-specificity threshold): Sensitivity 62.7% (59.9%–65.5%) and specificity 90.0% (89.4%–90.6%).
• Multi-task classification performance: Mean Top-1 accuracy 42.6±20.7%, Top-2 56.1±22.8%, and Top-3 61.9±22.9%.
• Clinical comparison (diagnoses by physicians): Clinical diagnoses achieved Top-1 65.4±17.7%, Top-2 73.9±16.6%, and Top-3 74.7±16.6%; physician Top-3 sensitivity 88.1% and specificity 83.8%.
• Reader test comparison: RCNN reader test results showed sensitivity 66.9%±30.2% and specificity 87.4% in comparative analyses.
• External validation: Performance was validated against diagnoses made by 65 attending physicians at the time of biopsy request.

Scientific Applications:

• Malignancy screening: Distinguishing benign versus malignant skin tumors from clinical photographs for diagnostic assessment.
• Dermatological diagnosis benchmarking: Comparing automated convolutional neural network performance against attending physician diagnoses for validation studies.
• Telemedicine and mass screening: Photograph-based screening applications for large-scale or remote dermatologic assessment.

Methodology:

Applied deep learning with convolutional neural networks to unprocessed clinical photographs; trained and evaluated on 40,331 images from 10,426 patients covering 43 disorders (Jan 1, 2008–Mar 31, 2019); performed binary and multi-task classification and externally validated performance against diagnoses by 65 attending physicians at biopsy request.