DeepDerm

• 

  DeepDerm

Inspiration

We were thinking of ways to use machine-learning and deep-learning to the advantage of healthcare, so that patients can seek care early, have an early diagnosis, with minimal invasiveness and can be started on treatment promptly. In the early stages of some types of skin diseases, even dermatologists (your skin doctor) find it difficult to differentiate between them because of the overlapping clinical features of these diseases. Even when biopsied, these diseases show similar histological features. How it is treated depends on the condition you have. The quicker you start treatment, the less likely you are to have severe forms and scars. So reaching an early diagnosis is of utmost importance. We used a dataset of six dermatoses namely psoriasis, seboreic dermatitis, lichen planus, pityriasis rosea, chronic dermatitis, pityriasis rubra pilaris (collectively called the “erythemato-squamous disorders” or ESD), to automatically detect the specific disease in someone who has similar symptoms. We applied machine-learning and deep learning techniques to this data set for disease prediction according to the attributes a particular patient has.

What it does

Inflamed skin is easy to spot but it is hard to figure out the cause. Which could it be? DeepDerm diagnoses your skin condition with a 98.6% accuracy so that your dermatologist can start the specialized treatment at the earliest, and you can have it under control in no time. DeepDerm automatically detects which of the “erythemato-squamous disorders” a patient has, to support physicians in a correct diagnosis, which is very important on the path leading to early treatment and good prognosis.

How we built it

We used Python and TensorFlow to develop the engine behind DeepDerm

Challenges we ran into

Our biggest challenge was to use the correct techniques to get the highest accuracy

Accomplishments that we're proud of

This method achieves 98.6% accuracy, which is currently the best method available from this data set according to literature. We were able to implement the DeepDerm engine within the time frame of the Hackathon!

What we learned

Applying deep learning techniques

What's next for DeepDerm

We can utilize “dermoscopy" images and train them to automatically diagnose all types of skin diseases including Melanoma, which is a skin cancer.

Dermoscopy, which is a newer and non-invasive modality for the diagnosis of skin diseases, utilizes a high quality magnifying lens and a powerful lighting system (a dermatoscope) which allows examination of skin structures and patterns which are then diagnosed by an expert physician.

And better yet, we can use a smartphone to do this, with a database of thousands of images, which is great for resource limited settings. It also has the potential to achieve a diagnosis without expert opinion.

Sensitivity and specificity of using this technique for diagnosis of skin disease can be tested and calculated before it is used as a diagnostic tool in the clinical setting.

Acknowledgements

The dermatology data set used in this application was obtained from the publicly available Machine Learning Repository of University of California Irvine.

Reference: Lichman, M. (2013). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science.

Built With

• ai
• deep-learning
• google-cloud
• machine-learning
• pandas
• python
• scikit-learn
• scipy
• tensorflow