The 99% to 25% Gap: Why Early Detection Matters

When I learned that melanoma survival rates drop from 99% to less than 25% when caught late, I knew I had to build something that could make early screening accessible.

Figure 1: Survival rates of melanoma based on stage. 2,10,11 (cited from https://www.researchgate.net/figure/Survival-rates-of-melanoma-based-on-stage-2-10-11_tbl3_328040775)

Skin cancer doesn't wait for specialist appointments. In rural communities, wait times can stretch months. In urban centers, dermatology visits are expensive and often overbooked. Meanwhile, suspicious moles go unchecked.

NOMA AI started as a question: Could we build a device that puts preliminary screening in anyone's hands? Not to replace doctors, but to empower people with information before it's too late.

This blog series documents the journey, from dataset collection to model training to hardware integration, of building an accessible skin cancer screening tool.

Key Challenges:

- Finding a diverse, comprehensive dataset (12,900 images across 24 conditions)

- Balancing class sizes (some cancers had 200 images while benign conditions had 1000+)

- Optimizing a model to run on a Raspberry Pi (MobileNetV3 was the answer)

- Making AI explainable (Grad-CAM heatmaps)

Read Part 2 to learn about the dataset and preprocessing strategy.