A research team has published a framework that classifies 14 chest conditions from X-rays — and, crucially, explains its reasoning in anatomically consistent terms doctors can audit.
The system, called PulmoSight-XAI, trains separately on frontal and lateral X-ray views using an ensemble of five convolutional neural networks. Rather than collapsing spatial detail with standard pooling, it uses attention modules that preserve fine-grained features at multiple scales. A two-level meta-learning stack — three gradient-boosting models at the first level, blending at the second — aggregates predictions and uncertainty estimates at inference time. On a CheXpert-style benchmark, the framework scored 0.9319 macro-average AUROC on frontal views and 0.9154 on lateral, both reported as state-of-the-art for the 14-condition multi-label task.
The "explainability" angle matters more than the benchmark number. Most clinical AI rejection comes not from low accuracy but from clinicians not trusting a black box — seven post-hoc attribution methods here try to show which image regions drove each prediction, giving radiologists a trail to follow or contest. That is the gap most chest X-ray models have failed to close.
Whether hospital procurement teams find a five-model ensemble practical at scale is a separate question the paper does not answer.