InlabHealth: A Collaborative Training Platform for Healthcare AI Models

The development of robust artificial intelligence (AI) models in healthcare frequently encounters barriers related to data silos and strict privacy regulations. To address these challenges, we introduce InlabHealth, a collaborative AI model training platform designed specifically to accelerate healthcare innovation. InlabHealth provides a collaborative ecosystem where users can create, share, and iterate on machine learning models. A core feature of the platform is its transparent training environment, which grants access to in-progress training states. This allows researchers and institutions to download isolated model weights or seamlessly participate in distributed training via federated learning (FL) techniques. The platform's architecture and practical viability were successfully validated through an international proof-of-concept. Utilising InlabHealth, universities in the United States and Brazil were able to asynchronously and collaboratively train a pulmonary nodule segmentation model on three-dimensional computed tomography (CT) imaging. This cross-border collaboration demonstrates the platform's efficacy in fostering decentralised, secure, and continuous medical AI development without requiring the direct sharing of sensitive patient data. Beyond demonstrating feasibility, our proof-of-concept produced a mechanistic finding relevant to medical federated learning. When the two site models are trained from independent random initialisations, the canonical parameter-averaging aggregator (FedAvg) collapses to a degenerate all-background predictor (Dice score 0.000 on every one of 60 held-out test volumes), because the linear path between the two site optima in weight space crosses a high-loss barrier between different basins of the loss landscape. This failure is fully explained by the absence of linear mode connectivity, and the platform implements two complementary remedies. The first is a shared-initialisation protocol that enforces a common starting point for every federation round and recovers FedAvg monotonically with the depth of the shared initialisation. The second is a Fisher-weighted merge aggregator that reweights each parameter by its empirical Fisher information and recovers a functional model (Dice ? 0.19) even from inputs that the naive average destroys. Both options are exposed by the platform so institutional partners with heterogeneous training schedules can still federate productively.