Truace
Subscribe

UBNet-Seg cardiomegaly screening accuracy on external NIH and OpenI datasets

Source article: AI-Assisted cardiomegaly screening via implicit morphological inference and human-in-the-loop validation

Cardiomegaly screening via manual Cardiothoracic Ratio (CTR) measurement remains a clinical bottleneck, while contemporary deep learning solutions often suffer from algorithmic bloating. To address the need for resource-efficient and interpretable triage, this study proposes a framework driven by implicit morphological inference, which bypasses the requirement for explicit heart segmentation. We developed UBNet-Seg, a lightweight U-Net variant (2.3 million parameters) trained on a heterogeneous dataset of 11,748…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-0257· cite & verify
AI-Assisted cardiomegaly screening via implicit morphological inference and human-in-the-loop validation
Cardio surgery 3 by Dominique Cappronnier. CC BY-SA 3.0 · https://creativecommons.org/licenses/by-sa/3.0
Trace impact readingContested
P 73The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 70The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

A peer-reviewed study published July 16, 2026 describes UBNet-Seg, a lightweight 2.3-million-parameter U-Net variant that infers cardiomegaly from lung field geometry rather than explicit heart segmentation. Trained on 11,748 images, it was evaluated on external NIH and OpenI chest X-ray datasets, reporting 95.85% lung Dice and 0.05-second inference.

The work matters because it demonstrates a low-latency, explainable triage alternative for resource-constrained settings, while showing that domain shift still reduces fully automated performance on OpenI to 76.07% before human-in-the-loop correction lifts it to 91.21%. Whether the proxy approach generalizes to broader clinical populations and real workflow remains to be tested prospectively.

Main points

  1. Developed UBNet-Seg, a lightweight U-Net variant (2.3 million parameters) trained on 11,748 images to segment lung fields as geometric proxy, bypassing explicit heart segmentation.
  2. External validation on NIH and OpenI showed lung Dice 95.85% and inference time 0.05 s, with fully automated accuracy 90.31% on NIH and 76.07% on OpenI.
  3. Human-in-the-Loop refinement neutralized domain shifts, with McNemar test confirming improvements to 93.63% and 91.21% (p<0.001).

The gain

UBNet-Seg, a 2.3M-parameter U-Net variant using lung fields as geometric proxy, achieved 95.85% lung Dice at 0.05s inference and 90.31% automated cardiomegaly accuracy on NIH, rising to 93.63% on NIH and 91.21% on OpenI after expert-guided refinement.

The problem

Fully automated cardiomegaly screening accuracy dropped to 76.07% on the external OpenI dataset, showing domain-shift vulnerability, while manual CTR measurement remains a clinical bottleneck and existing deep models suffer from algorithmic bloating.

The rundown

Researchers trained UBNet-Seg, a 2.3 million parameter U-Net variant, on 11,748 heterogeneous chest images to segment lung fields as a proxy for heart size, implementing implicit morphological inference to avoid explicit heart segmentation.

Tested on unseen NIH and OpenI datasets, the model reported 95.85% lung Dice and 0.05 s inference, with automated accuracy of 90.31% and 76.07%; adding expert-guided refinement raised accuracy to 93.63% and 91.21% respectively, with p<0.001 on McNemar test.

What this doesn’t fix

Fully automated performance degraded substantially on external OpenI data, requiring expert-guided refinement to restore accuracy, indicating residual domain-shift vulnerability.

Sources

  1. Peer-reviewedJournal of X-Ray Science and Technology2026-07-16

The debate