Abstract
The development of robust artificial intelligence models in healthcare frequently encounters barriers related to data silos and strict privacy regulations. To address these challenges we introduce InlabHealth, a collaborative training platform on which researchers and clinical institutions can create, share, and iterate on machine-learning models. A core feature of the platform is its transparent training environment, which grants every project participant access to in-progress training states and to the resulting model weights, so that partner sites can either deploy a chosen checkpoint directly or seamlessly participate in distributed training rounds using federated learning.
The platform's architecture and practical viability were validated through an international proof-of-concept. Using InlabHealth, universities in the United States and Brazil asynchronously and collaboratively trained a three-dimensional pulmonary nodule segmentation model on chest computed-tomography scans drawn from a publicly available archive partitioned into two disjoint patient cohorts, one held at each site. This cross-border collaboration demonstrates the platform's efficacy in fostering decentralised, secure, and continuous medical artificial intelligence development without requiring direct sharing of sensitive patient data, on top of the heterogeneous hardware that real-world hospital partnerships typically present.
Our experiment produced a mechanistic finding relevant to medical federated learning. With independent random initialisations, naive parameter averaging collapsed to a degenerate all-background predictor (mean Dice zero on all sixty held-out test volumes), because the linear path between the two site optima crossed a high-loss barrier between different basins of the loss landscape. This failure is fully explained by the absence of linear mode connectivity. The platform implements two complementary remedies. A shared-initialisation protocol enforcing a common starting point recovered the parameter average monotonically with the depth of the shared initialisation, and a Fisher-weighted merge that reweights each parameter by its empirical curvature recovers a functional model (mean Dice approximately 0.19) even from inputs that the naive average destroys. A sequential warm-start protocol that finetunes the partner site's model from the originating checkpoint reaches a mean Dice of 0.22 on the partner's test volumes — the deployment-realistic upper bound recommended by this work.