The Problem and our Solution

The Problem and our Solution

People routinely dismiss skin abnormalities as being harmful due to their rather unexceptional presentation. Thing is, while most skin cancers may not kill a person, they can substantially affect a person’s quality of life. Pain, infections, and disfigurement are all textbook consequences, and some conditions, like melanoma, have the capacity to spread to other organs and be deadly.

The first signs of a serious skin cancer typically present in very unspectacular ways. Moles, skin tags, and rashes are rarely viewed as causes for concern and are often dismissed as a threat until the discomfort is greater than the hassle of seeing a doctor. But it is during that time of latency where dangerous conditions can fester and become deadly.

When people finally take action, they are then subject to long wait times or need to travel hundreds (if not thousands) of kilometres to access physicians in a major urban centre.

Skinopathy is developing ground-breaking Artificial Intelligence (AI) and Augmented Reality (AR) technology that will not only give Canadians unparalleled access to physicians, online and in a matter of hours, but also advance medical practices by:

  • Creating an online platform and digitizing the doctor’s office.
  • Drastically reducing the time to see a physician from months to hours.
  • Integrating the service with healthcare insurance plans.
  • Using AI to triage patient urgency.
  • Informing diagnoses.
  • Educating patients about diseases.
  • Guiding surgeries through Augmented Reality (AR).
  • Preventing recurrence through regularly scheduled screening

We are also planning to expand to other skin abnormalities, including wounds, burns, and other dermatological diseases. Once ready, our technology will have the potential to help millions, if not billions, of people worldwide. Below we go into details about our technology, methodology, and reasoning.

Please do not hesitate to contact us if you have any additional questions or wish to try the application yourself.

The Technology

The Technology

There are several distinct types of artificial neural networks. Some are created by connecting dense layers on top of dense layers, others are used to break large tasks into smaller pieces, and some are even used for predicting a time-series event.

What we have created is a Convolutional Neural Network (CNN)-based technology geared for skin abnormalities. CNN’s are mostly used for analysing images. The most famous application being the “Re-Captcha” security feature found on many websites, which is a process that trains such CNNs.

However, instead of using our AI to determine the difference between a fire hydrant and a bus, we are using this technology to determine the difference between a mole and a cancerous lesion.

The technology is also being future-proofed for other use-cases, which is why we have two development schemas:

  • Narrow – used solely for classifying skin abnormalities. This proprietary light-load AI will allow for super accurate identification of skin abnormalities while being less computationally intensive.
  • Broad – used for classifying everyday objects. This will be a computationally-expensive but generalized CNN, whose code will be open-sourced at a later date. This will allow non-commercial researchers, or licensed commercial users, to adapt our AI model’s architecture and optimization strategies for other specific use cases.

Our approach was first to determine both the viability and accuracy of the Skinopathy concept using a simple architecture before we began creating SOTA technology. That is why we began testing our concept using a pre-trained model with DenseNet, a slightly older AI architecture that made it easy for us to quickly test a small-scale prototype.

Manish Chablani, Head of AI and Research at Eight Sleep, describes DenseNet this way:

[DenseNet] connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections — one between each layer and its subsequent layer — our network has L(L+1)/ 2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters.

HAM10000 vs Real World

HAM10000 vs Real World

To date, we used two databases to test our technology. First was the HAM10000 database, a large and thoroughly curated open-source database with over 10,000 confirmed images of various skin abnormalities. This database was built by physicians who used specialized tools, like dermatoscopes, to take the perfect image of a skin abnormality. The HAM10000 database is generally regarded as the gold standard for training any skin related AI architecture.

However, our technology is geared for the public. Not trained physicians with dermatoscopes. That is why we created a secondary database that more accurately reflected the type of real-world images patients are likely to submit. The database we built came from patients who used their cellphone and then kindly submitted those images to our AI team.

The substantial variance in both focus and field of view allowed us to do real-world testing of our model and compare those results with the pristine HAM10000 database.

A Better Mouse Trap

A Better Mouse Trap

Testing results from our small-scale prototype were very encouraging. The categorical accuracy of our neural network yielded 92% accuracy using the HAM10000 database and 83% with our own proprietary database (~400 images).

Now that we had confirmed the small-scale prototype, we were ready to move onto a more sophisticated model. We chose EfficientNet B7, the latest version of the Google open-source AI architecture that was released in the fall of 2020. The abstract for the original paper explaining EfficientNet can be read below:

Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance.

Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet.

Simply said, Google created a better, more efficient, and open source mousetrap.

The Tensorflow-Keras package used by EfficientNet B7 allowed the kind of flexibility we needed to potentially scale our technology in the future. It is also SOTA and we wanted to ensure our concept could withstand its rigour before we began developing our own proprietary SOTA for skin abnormalities.

We once again tested the technology using the HAM10000 database and our own proprietary images, and once again we were incredibly happy with the results. The HAM10000 databases yielded 90% accuracy while our own proprietary images yielded 87% accuracy.

That 4% increase from DenseNet to EfficientNet B7 with our images was the most encouraging because it confirmed our hypothesis: A more sophisticated AI architecture geared for one simple task could take real-world camera images, with a high variance in image quality, and still produce highly accurate results.

Patent Pending

Patent Pending

We have now built technology and filed a patent with the USPTO that features unique Intellectual Property that will make our technology particularly impactful, including:

  • An integrated feedback loop process that continuously refines our precision and accuracy.
  • A light-load AI that dramatically reduces computational power requirements.
  • Combining AI and AR to offer physicians precision excision capabilities.
  • And innovative tools that expand our capabilities well beyond skin cancer.

The Canadian Skin Cancer Foundation states that 1 in 3 cancers diagnosed worldwide is skin cancer and that they outnumber lung, breast, prostate, and colon cancers combined.

Since the technology is geographically agnostic, we will be able to be deploy it to under-serviced regions and provide people with unparalleled access to healthcare. No more flights. No more lodging arrangements. And no more waiting for the skin cancer specialist to visit your town.

Share This