Description

SkinDx is Machine Learning freeware mobile app for pigmented skin lesion diagnosis.

This app uses Tensorflow and Tensorflow Lite for developing machine learning mobile app. The Mobilenet v1 quantized model is used for transfer training with dataset from ISIC (benign vs malignant) and HM10000 dataset. The accuracy test of both models are evaluated on mobile by unseen validating datasets. The scores are as follow:

ISIC (Benign vs Malignant) dataset

  • benign sensitivity: 75.00% (135/(135+45))
  • benign precision: 87.66% (135/(135+19))
  • malignant sensitivity: 87.25% (130/(130+19))
  • malignant precision: 74.29% (130/(130+45))
  • Total accuracy: 80.55% (265/329)

HAM10000 dataset

  • akiec sensitivity: 25.00% (8/(8+24))
  • akiec precision: 57.14% (8/(8+6))
  • bcc sensitivity: 35.29% (18/(18+33))
  • bcc precision: 72.00% (18/(18+7))
  • mel sensitivity: 47.75% (53/(53+58))
  • mel precision: 27.75% (53/(53+138))
  • bkl sensitivity: 25.69% (28/(28+81))
  • bkl precision: 45.90% (28/(28+33))
  • nv sensitivity: 87.46% (586/(586+84))
  • nv precision: 85.30% (586/(586+101))
  • vasc sensitivity: 14.29% (2/(2+12))
  • vasc precision: 28.57% (2/(2+5))
  • df sensitivity: 45.45% (5/(5+6))
  • df precision: 38.46% (5/(5+8))
  • Total accuracy: 70.14% (700/998)

*Sensitivity calculated by True positive / ( True positive + False negative)

*Precision or Positive Predictive Value (PPV) calculated by True positive / ( True positive + False positive)

You can use this app only for screening skin lesion. It CAN'T be used for diagnosis without any supervision from medical expert.

Screenshots

Video

-

Platform requirement

  • Android 6.0 or later

History

  • Version 1.2.0 released (25 July 2019)
    • Quantized model
  • Version 1.1.0 released (23 July 2019)
    • Model improvement
    • Add model accuracy test
    • Add model loading and resetting
  • Version 1.0 released (8 July 2019)
Get it on Google Play