DermaScan

• 

  Let AI detect the signs early.

Inspiration

The app DermaScan is for detecting skin lesion type from comfort of one's home. Skin cancer is on of the most common cancers in the world. Early detection of signs of skin cancer can increase the chances for proper treatment and recovery to a great extent. Therefore, I wanted to build something that would be user friendly, immediate result and needs nothing but a smart phone to use.

What it does

The app has very simple working process. In the home page along with basic information regarding skin cancer there is option for scanning. There the user is asked to drag and drop an image of the lesion. Within seconds, the probable category of lesion is shown below. Often the lesion shows signs of more than one category. In that case, the result is shown as percentage of multiple categories. The datast used to build this app HAM10000 had 7 categories of skin lesion images- Actinic keratoses and intraepithelial carcinoma (akiec), Basal cell carcinoma (bcc), Benign keratosis-like lesions (bkl), Dermatofibroma (df), Melanoma (mel), Melanocytic nevi (nv), Vascular lesions (vasc). There is segment in the app under the name Education where details of these types of lesions, their risk, causes etc. is presented.

How we built it

HAM10000 is open access dataset. We used this dataset to train our model. The app is built completely on bolt.new and the prompts were structured by ChatGPT. Afterwards the app is deployed using Netlify.

Challenges we ran into

The biggest challenge was ensuring the accuracy of the results. My priority was to avoid false negative at all cost. The overall accuracy of the model is around 88%.

Accomplishments that we're proud of

The app can be stepping stone for future skin cancer detectors which will be built on bigger database, will be robust & better. Moreover, the same knowledge distillation architecture can be used for other cancer detection model given dataset of that particular cancer images.

What we learned

Different knowledge distillation process were learnt while building the model. Bolt.new is a great platform to build apps for free. In this way our works can get a usable form and accessible to public as well.

What's next for DermaScan

The dataset of HAM10000 is fair skin biased. The next step is to use GAN to generate lesion images on different tone of skin and developing the model further so that anyone can use the app and the accuracy of result doesn't fall.

Built With

• bolt.new
• english
• ham10000
• javascript
• netlify
• pytorch