Model Dermatology with RCNN

Model Dermatology with RCNN DEMO (ver.201907)

Check status...

Drop image here
or click to upload.

This is an experimental DEMO for "Keratinocytic Skin Cancer Detection on the Face using Region-based Convolutional Neural Networks" (JAMA Dermatology) and "Assessment of Deep Neural Networks for the Diagnosis of Benign and Malignant Skin Neoplasms in Comparison with Dermatologists: A Retrospective Validation Study". The submitted image will be transferred with IP address and stored as a database, which will be used only for the development of the algorithm. It will take 5~10 seconds to upload an image and additional 5~10 seconds to analyze one image.

Instruction and Warning
1. We strongly recommend to use high resolution images over 3000x2000 pixels. Currently, the minimal resolution of the blob detection is set to 256x256 pixels.
2. Please test with two macro (extreme close-up) and two close-up (distance enough to fill face in the frame) images with different angles. Select the final value as the highest value among the results of four photographs.
3. The secondary changes of the lesion (i.e. laser surgery, curretage) may be a cause of false positive.
4. The diagnostic performance made with images alone was inferior to that of actual practice.

The Result of the Retrospective Study
- Algorithm
Red rectangle : Sensitivity = 79.1%, Specificity = 76.9%
Orange rectangle : Sensitivity = 62.7%, Specificity = 90.0%
- Actual Practice in a Tertiary Hospital
From Top-3 : Sensitivity = 88.1%, Specificity = 83.8%
From Top-1 : Sensitivity = 70.2%, Specificity = 95.6%


New Model (build 2020.08)
New improved model is available at The sensitivity using the ISIC dataset (; TAG = JID2018 Editorial Image) is improved from 52% to 70%.