Logo Lanfrica

vic-zzy/mira

Domaine:

healthcare

Type de record:

software
Créateur:
vic
Hôte:
Cervical cancer screening on a sub-$100 Android, fully offline, powered by Gemma 4. For low-resource clinics # Mira *An on-device cervical cancer screening assistant for low-resource clinics. Fully offline. Sub-$100 Android target. Gemma 4 powers multilingual patient communication, clinical decision support, and workflow orchestration on top of a domain-specific image classifier.* Built for the Gemma 4 Good Hackathon. --- ## The problem Cervical cancer kills around 350,000 women every year, and roughly 90% of those deaths happen in low- and middle-income countries. It is also one of the most preventable cancers, if precancerous lesions are caught early. The screening method recommended by the WHO for resource-limited settings is Visual Inspection with Acetic Acid (VIA): swab the cervix with diluted vinegar, look for tissue that turns white, refer if needed. Cheap, fast, no lab required. The bottleneck is the trained eye. Community health workers in rural clinics learn VIA in short courses and often lack the experience to be confident. Specialists are scarce and concentrated in cities. Patients get referred too often or not often enough. Mira is built around that gap. ## What Mira does The flow is short. A community health worker (CHW) prepares the patient, waits 60 seconds after applying acetic acid, and taps capture. From that tap: 1. The phone fires three photos in a burst, drops any that are blurred by motion, and averages the rest into one composite. 2. A content validity gate refuses obviously non-cervical inputs. 3. An on-device classifier produces a probability and a 16×16 saliency map showing where the model looked. 4. Gemma 4, running locally on the phone, takes the result, the saliency map, and the patient context, and produces a multilingual explanation, suggested next steps, and structured workflow actions. 5. The CHW reads the suggestion, listens to the explanation in the patient's language, and makes the clinical call. Everything runs on the device. No internet. No images leave the phone. The app ships in ten languages: English, Spanish, Portuguese, …