TinyTB-Net: Building an Offline Tuberculosis Screening Model for a $2.50 ESP32 Microcontroller

An engineering write-up describes building TinyTB-Net, a compact vision model trained from scratch to screen chest X-rays for tuberculosis and run on a roughly $2.50 ESP32 microcontroller without internet access. The author frames the project against the roughly 1.3 million annual TB deaths, noting that diagnosis rather than treatment is the main bottleneck, especially in regions with very few radiologists. The piece stresses that plain accuracy is misleading for medical triage, since missed infections (false negatives) carry far higher costs than unnecessary follow-up tests (false positives). Instead, the author emphasizes sensitivity and related metrics that reflect those asymmetric error costs. The article also covers architectural choices, INT8 quantization, and serving a web interface from microcontroller flash. It is presented as an educational build, not a regulated medical device.