Spotter uses computer vision to predict if freckles or moles have a high probability of melanoma
In 2016, ten thousand people will likely die of melanoma. In countries such as the United States, many if not the majority of these deaths could have been prevented.
According to skincancer.org, "The estimated 5-year survival rate for patients whose melanoma is detected early is about 98 percent in the U.S. The survival rate falls to 63 percent when the disease reaches the lymph nodes, and 17 percent when the disease metastasizes to distant organs."
As I, an individual with a voluminous amount of freckles and moles, I must get a yearly body exam from a doctor to ensure of no skin irregularities. Many, however, can not afford these exams, which often cost around $150 to $200.
With these ideas in mind, we developed a website to help individuals detect malignant melanoma and atypical skin lesions on their persons. Users submit an image and within seconds that image is analyzed for its probability of containing malignant melanoma.
Using Clarifai's API, I created two sets of images. One set contained only malignant freckles or moles. The other set contained only benign freckles or moles. With these two datasets, I was able to input a new image and compare it to both image sets. Clarifai is able to understand that there are clear differences between the two sets as severely malignant moles have clear physical peculiarities such as being, for example, asymmetrical and multicolored.
Then I parsed the probability of an image being malignant and displayed it. Moreover, I analyzed the probabilities I received and created a scale that I believe is impressively accurate.
The Scale: The probability is on a scale from 0% being the most benign to 100% being the most malignant. Through my own research and modeling, I used the following logic.
Probability of Malignancy:
- High: Greater than 70%
- Moderate to High: 66% to 70%
- Low to moderate: between 56% and 66%
- Low: less or equal to 56%
- Not a freckle: less than 20%
I found this scale to be very impressive. While I am in no way suggesting it meets the standards to be used medically nor do I believe it should be used for anyone to make any form of medical decision, I was highly proud of myself for creating this piece of software, that at least in theory, could save someone's life.
The biggest challenge I had with the project had to do with user input and saving images. I used heroku and thought that I would be able to skip using a proper database of images in favor of folders of images. This turned out to be naive and I wasted several hours by trying for this shortcut, ultimately deciding to use Amazon Web Services. The second biggest challenge I had was just about building my dataset to be as accurate as possible.
In the future, I would like to make my dataset have several hundred images in order to be more scientifically sound.