Nov, 2019

The application of deep learning model in image classification and feature detection is becoming widespread in the medical community. This webpage presents a deep learning model that takes a dermascopic image along with meta data to infer the type of lesion, and it has resulted in performance improvement compared to the base model without the metadata. Please note this is NOT a diagnostic tool, the output of the model is not to be interpreted as an actual diagnosis. In practice, such a platform would serve the purpose of monitoring skin lesions (especially for patients with family history) and encouraging the user to seek medical attention more efficiently.

A DenseNet model was trained on data set provided by the International Skin Imaging Collaboration (ISIC). The goal was to classify dermoscopic images for 8 categories (the bolded ones are cancerous):

  1. Actinic Keratosis
  2. Basal cell carcinoma
  3. Benign Keratosis
  4. Dermatofibroma
  5. Melanoma
  6. Nevus (mole)
  7. Squamous cell carcinoma
  8. Vascular lesion

The ISIC 2019 data is provided courtesy of the following sources: BCN_20000 Dataset: (c) Department of Dermatology, Hospital ClĂ­nic de Barcelona HAM10000 Dataset: (c) by ViDIR Group, Department of Dermatology, Medical University of Vienna; MSK Dataset: (c) Anonymous; ;

Fig.1. Sample images from the ISIC archive.

With the incorporation of data for gender and age, the model performance has been improved and the following table shows the metrics of the model when evaluated on the validation data set.

Table 1. Performance metrics of the DenseNet classifier (with meta data)
Category Precision Recall F1 AUC
Actinic keratosis 0.57 0.67 0.62 0.97
Basal cell carcinoma 0.84 0.84 0.84 0.98
Benign keratosis 0.67 0.76 0.71 0.96
Dermatofibroma 0.55 0.92 0.69 0.99
Melanoma 0.76 0.75 0.76 0.95
Nevus 0.92 0.88 0.90 0.96
Squamous cell carcinoma 0.57 0.74 0.65 0.98
Vascular lesion 0.86 0.90 0.88 0.72

How to use

To run the app, simply upload a dermascopic image and enter information for gender and age below. All the information are required for the model to make its prediction.

To help you get started, feel free to use these sample images from ISIC.

ISIC_0028537(Melanoma, age 35, Male), ISIC_0030207(Benign keratosis, age 75, Female), ISIC_0072880(Basal cell carcinoma, age 85, Male), ISIC_0058863(Melanoma, age 45, Female), ISIC_0058864(Actinic keratosis, age 70, Male), ISIC_0060302(Squamous cell carcinoma, age 45, Female), ISIC_0069896(Nevus, age 50, Male), ISIC_0034094(Melanoma, age 55, Female), ISIC_0055686(Actinic keratosis, age 85, Female)


For questions and comments please contact Jenny Yu (jypucca[at]hotmail[dot]com)