Melanocytic lesions suite
This suite is a blend of advanced AI models and algorithms, each contributing to a nuanced understanding of melanocytic lesions.
Components and Capabilities
Melanoma vs. Nevus Classifier employs sophisticated AI-algorithms to distinguish between cutaneous melanomas and nevi.
The Melanoma Segmentation Model localizes tumor regions on a slide.
The Ulceration Classifier leverages advanced machine learning techniques to accurately classify ulcerations in melanoma cases.
The Tumor Depth and Diameter Estimator automatically measures the primary tumor.
Dataset Used and Metrics
Melanoma vs. Nevus Classifier
The model was trained on a private dataset comprising 9,196 whole slide images (WSIs), which include 3,290 melanomas and 5,906 nevi. Professional dermatopathologists annotated each slide, ensuring the highest quality of annotation.
Internal validation dataset of 2323 WSIs, with 801 melanomas and 1,520 nevi:
Accuracy: 0.93
Sensitivity: 0.94
Specificity: 0.93
Another internal dataset of 14,133 nevi WSIs:
Accuracy: 0.94
Publicly available TCGA dataset of 948 melanoma WSIs:
Accuracy: 0.96
Publicly available CPTAC dataset of 411 melanoma WSIs:
Accuracy: 0.93
Ulceration Classifier
The model was trained on 2,632 melanomas, annotated by professional dermatopathologists. On our internal validation dataset consisting of 658 WSIs this model demonstrates:
Accuracy: 0.91
Sensitivity: 0.92
Specificity: 0.91
Tumor Segmentation Model
Trained on a dataset of 1,213 WSIs containing tumors, with segmentation masks created by professional dermatopathologists, the model achieves:
IoU: 0.89
Accuracy: 0.95
Sensitivity: 0.92
Specificity: 0.98
Precision: 0.71
Recall (Sensitivity): 0.89
F1 Score: 0.79
Technical Stuff
This widget works with most modern web browsers and mobile devices. Just a heads-up, having at least 2GB RAM and a stable internet connection will make things run super smoothly.