☆ cancHER ☆

Empowering and leading women on their road to health and success.

Comment 6

INSPIRATION 💭💝

I want to create something, I thought, I wanted to create something impactful. Staring at my laptop screen with a blank mind, I wasn't entirely sure what to do at the beginning of this hackathon, but then I thought of my grandfather and wondered what he would have told me. I remembered how much he wanted me to find my passion and work hard to achieve my goals. Even though it was a distant memory, I felt a sudden spark of inspiration and wanted to pursue something I never thought I would have been able to do a week ago: create a cancer detection model. I wanted to show him that I am trying my best to combat such a widespread issue that potentially left families emotionally and physically suffering for years. I have a feeling he would have approved. 🎗

FUNCTIONALITY 💻

This model is very portable because it can be downloaded into one's drive through colab features, so implementing it into a product's biosensors seems to be a practical solution. But diving deeper into the model itself, its features include being able to classify breast cancer (cell) images into categories through neural networks and how the model(s) was/were trained. (For testing purposes, there were three models, which are explained more in the next section.) The biosensors would be able to detect any abnormal lumps or cell divisions in the breast area and notify the patient of any changes.

CREATION PROCESS 🤔

The coding language I used was Python, and for my platform, I used Google Colab.

Using my background knowledge from an AI/machine learning summer camp I had attended, I created an artificial intelligence model (which uses machine learning) to classify whether or not the patient has breast cancer. I trained it on pre-made/public datasets (training data - 80%) and then used testing data (20%) to see its accuracy results and confusion matrices. I first used the k-nearest-neightbors concept to create a KNN model but realized its accuracy was only around 20%, which is not very reliable. Then, I used a CNN model without applying transfer learning, but its accuracy was also not great: only 26%. Finally, I ran a CNN model with transfer learning and its model accuracy was a stunning 69% and AUC ROC score was around a beautiful 92%.

Lastly, I used Krita to draw the bra prototype and Canva to design/pull the two drawings together with a nice pink background.

CHALLENGES 💪

While google colab was very organized and easy to write/read, its execution was much slower than I would have liked. In fact, before recording my demo video, I had to wait around twenty minutes just to run everything! 😂 However, colab still did work very well for each code block and its separate execution. There were also many errors that popped up, which is inevitable in CS, but the number of errors plus tiring runtime did take me a very long time to fully finish the program.

ACCOMPLISHMENTS + TAKEAWAYS 🥳🏆

I'm really proud of myself for pushing outside of my comfort zone because this program taught me how to use AI/machine learning and maneuver through different novel python libraries. Moreover, I felt that my time management (as well as networking skills!) significantly improved. Despite the errors, I didn't give up on this model because I knew I had that passion that would take me to the finish line. 🙂

FUTURE PLANS 🌏

Despite its successes, the detection model still has many limitations, including preventing AI bias/AI ethical concerns, whether skin color would affect the model, runtime issues, practicality, etc. However, there are a few solutions that I can think of without too much time, such as just training the model with a more diverse range of skin color pictures/skin textures to prevent racial bias. I believe that as long I persevere through these hardships, I can create a program that actually changes lives and impacts families. 💝

Built With

Share this project:

Updates