Guppy - Cancer Detecting A.I.

Our logo

Inspiration

Obstacle 1: The Hardware Crash We tried to run the training on Rishav's old laptop. It got so hot we could literally cook an egg on it, and then it crashed. We lost two days of work.

Solution: We realized we couldn't use a CPU. We moved our code to Google Colab to use their cloud GPUs (Graphic Processing Units), which are way faster at doing matrix multiplication.

Obstacle 2: The "Overfitting" Trap At one point, Advait shouted, "We got 100% accuracy!" We were celebrating until we tested it on a new photo, and it failed completely. The AI had memorized the specific training photos instead of learning the patterns. It was like memorizing the answers to a practice test but failing the real exam.

Solution: We added a "Dropout" layer. This randomly turns off some neurons in the brain during training, forcing the AI to not rely too much on any single pixel.

Obstacle 3: Imbalanced Data We had way more photos of healthy skin than cancer. The AI became lazy and just guessed "Healthy" every time because it was right 90% of the time by luck.

Solution: We used Data Augmentation.

We wrote code to flip, rotate, and zoom the cancer images to create "fake" new examples. This balanced the dataset so the AI had to actually learn.

The Result After 25 epochs it reached 67% accuracy and 84% sensitivity.