The problem EARLY SKIN DISEASE DETECTION solves

** ABSTRACT **:

• **Today skin diseases are one of the most common diseases, whose number has been increasing day by day. Mostly, people neglect skin diseases and treatment procedures.This neglectance at the initial stages proves pernicious. But even if people consult physicians, it is quite difficult for them to precisely detect skin diseases with high accuracy and precision. Especially when it comes to the diseases like Melanoma, Actinic Keratoses, Basal cell carcinoma, Benign keratosis, Dermatofibroma, Melanocytic nevi, etc. **

• ** Predictions of skin diseases in their early stages is today's need. For this, we have built a multi-layered Early Skin Disease Detection MobileNet CNN model(proposed as a deep learning model by Andrew Howard) having 91 layers in it. **

• ** We have used this model to classify images in following 9 classes: **

1. Actinic Keratoses
2. Basal cell carcinoma
3. Benign keratosis
4. Dermatofibroma
5. Melanoma
6. Melanocytic nevi
7. Vascular skin lesions
8. Normal Human Skin Patch
9. Un-Zoomed human images
   

• ** The dataset for above 9 classes is prepared from following sources: **

  1. ** HAM-10000 for:**

     • Actinic Keratoses
     • Basal cell carcinoma
     • Benign keratosis
     • Dermatofibroma
     • Melanoma
     • Melanocytic nevi
     • Vascular skin lesions

  2. ** ISIC 2019 for: **

     • Actinic-Keratoses
     • Basal cell carcinoma
     • Benign keratosis
     • Dermatofibroma
     • Melanoma
     • Vascular skin lesions

  3. ** UTK-Face:** For Un-Zoomed human images

  4. Normal Human Skin Patch images were custom datasets built using Un-Zoomed human images. We cropped these images around forehead, cheeks and throat of human babies.

Challenges I ran into

** CHALLENGES:**

• ** Data Preparation: **
  • As we all know, data preparation in any ML project is the most cumbersome part. Extracting and cleaning data from different sources and integrating them is a bit tricky.
  • I prepared data from mainly 4 different sources given here :
    • ISIC 2019
    • HAM_10000
    • PH2Dataset
    • UTK-Face
• ** Classification: **
  • Skin rashes of Melanoma and Melanocytic-nevi are almost similar. Differentiating between them using even through human conscious is very difficult.

• ** Storage: **
  • Storing such a big model also requires a lot of space.

** SOLUTIONS**

• Instead of collecting normal human skin images from web. I used face images of UTK-Face Dataset and cropped it to create a custom skin dataset.
  

• I used the simplest approach for training(not requiring any CSV file). I prepared subclasses in the training and validation directory on the name of my classes and directly used them for model training.

• I removed majority of non-pigmented, subungual, ocular Melanoma which are similar to Melanocytic-nevi, to make Melanoma more differentiable.

• Also we have stored the model in .json, h5 format uploaded on drive and GitHub for further use