This is an EXPERIMENTAL DEMO APP for Medical Professionals. This DEMO produce a sub-optimal performance for the images of non-medical professionals. The results obtained with this program are for research purpose only.
** THIS DEMO CANNOT CONFIRM THAT THE SUBMITTED IMAGE IS NOT A SKIN CANCER. Approximately 10% of skin cancer can be missed if diagnosis is made based on clinical photography alone. Using the Edinburgh dataset, the model showed 90.4% sensitivity and 89.9% specificity for malignancy prediction at the high-sensitive threshold (cf) 80.3% sensitivity and 96.5% specificity at the specific threshold). The Top-1, 3, and 5 accuracy of this algorithm were 60.1%, 81.7%, and 88.7%.
** Please be sure to take pictures in bright places.
** The photographs used for analysis are transferred, and stored and can be used as research data for new service research and development.
– Top 1,2,3 Diagnosis
The app shows Top-3 diagnosis with score (range : 0.0~1.0). If the score exceeds 0.2, the diagnosis can be considered as a meaningful differential diagnosis.
– Malignancy Prediction
The algorithm displays the description of the chance of malignancy (miss rate = 1-sensitivity; range 0~100) as follows:
“Low” : miss rate – 0 ~ 10%
“Medium“, Biopsy recommended : miss rate – 10 ~ 20%
“High“, Biopsy recommended : miss rate – 20 ~ 100%
– Treatment Prediction
The treatment prediction is an experimental feature. If the score exceeds 50, you can consider the suggested options.
– Image Quality and Composition Assessment
The focus and brightness of submitted image crops are assessed by an adjective algorithm named “Fine Image Selector“. The algorithm evaluates composition of the image whether nodule or not, and the high composition score indicates that the image contains a nodule with proper composition. The Fine Image Selector also determines whether the submitted image is clinical image or general object.