Preprints.org
Distilling Multimodal Biomedical Foundation Models for Edge Tuberculosis Screening: Client-Side WebAssembly Deployment and Conformal Clinical Abstention
By Aditya MF.
Abstract
Background
Deep learning for tuberculosis triage risks shortcut learning and overconfidence. We evaluate distilling BioMedCLIP into an edge-deployable architecture with a selective conformal safety interlock.
Methods
A DenseNet-121 student (7.74M parameters) was trained on 12,599 radiographs via Tri-Loss distillation from a frozen BioMedCLIP teacher (86.4M parameters). Validation evaluated 3,441 radiographs: a primary within-scanner clinical benchmark (Combined India: Solan N=155, NITRD N=278) and an adversarial compression stress-test (Pakistan N=3,008). Mondrian split-conformal prediction calibrated a three-tiered interlock: urgent molecular testing, autonomous clearance, and clinical referral.