The Issue
- Utah has the highest rate of skin cancer in the country [citation].
- If it’s caught early, melanoma, is no big deal; however, if the disease progresses or spreads, the prognosis is notably worse. Consequently, prompt screenings are vital for effective treatment–treatments at stage 0 are much simpler, more pleasant, and notably cheaper.
- Who is well positioned to identify melanoma? Dermatologists.
The problem being: getting a derm appointment is an everloving pain.
Anecdote
Several years ago, I looked for a derm appointment at my neighborhood clinic, only to learn that none of the dermatologists were accepting new patients. I inquired at the next closest clinic, and schduled an appointment for two months later.
Dermatologists are in low supply and high demand.
It’s easy to imagine how patients could be discouraged by such wait times, not to mention issues of distance; potentially leading to substantial delays in care for undiagnosed melanomas.
Background
In the early 2020’s, I worked with a surgeon at Intermountain Healthcare to develop a curriculum for non-dermatologists to identify malignant melanoma lesions in collaboration with a professor from BYU. My role in the project was origially limited to stripping metadata from images sourced by Intermountain Healthcare, and sharing them via SFTP.
This project was supplemental and complementary to that project.
Structural Issues
There are entirely too many barriers to increasing the supply of physicians, much less dermatologists.
Urban-rural distribution of specialists futher exacerbate the issue in a large western state such as Utah.
Potential Solution
General practitioners or family physicians could use an appropriately trained convolutional neural network to refer patients to dermatologists expiditiously. GP’s and family doctors are unfortunately the target of most high-level initiatives–they are the doctors most likely to receive missives from… most everyone.
“See more patients in the same number of hours!”
“Encourage high-risk patients to get screenings.”
etc.
This is designed to provide a means of identifying patients with high-risk lesions and potentially referring them to dermatologists promptly.
Data
This project is designed to accept an in image of a skin lesion, i.e. a mole, and return a probability of that lesion being malignant. The data was sourced from kaggle. Computing resources were provided by Google Colab.
The Model
I made use of EfficientNetB0 for the purposes of transfer learning, trained the model, and briefly explored fine-tuning. There’s some evidence of the model effectiveness, but the heavy imbalance creates issues in evaluation.
Is this all?
That is to say, is the project complete? No.
If you examine the images, you’ll notice that:
- the images are remarkably well centered,
- most are well-lit, and
- most are on caucasian skin
If one wants to pass any random photo of a skin lesion, and have it evaluated accurately, a YOLO algorithm could help ameliorate the first issue. Similarly, anomaly detection might be able to help identify photos with poor lighting. What about non-caucasian skin?–I don’t have a good answer short of futher data collection.
Tools I used
|
|
|
|
Tensorflow is one of the preeminent packages for estimating neural networks, making GPU acceleration terribly easy. |
I estimated the model in Google Colabs. It provides easy access to cloud-based notebooks, and provides temporary free access to GPU's |
If you're looking to get started in machine learning, Kaggle is a good place to start. I made use of the data it makes publicly available. |
Conclusion
There is some promise to the approach, but I expect any further work will be dictated by the necessities of the work.