Diagnostic Loss introduces asymmetric sequence supervision for faithful radiology report generation

By pairing noisy-OR pooling for supported findings with log-sum-exp suppression for unsupported alternatives, Diagnostic Loss eliminates spurious hallucinations without altering model architectures.

Primary source: arXiv:2610.30980: https://arxiv.org/abs/2610.30980

What’s new: Automated radiology report generation requires vision-language models to produce accurate clinical descriptions while avoiding both omitted pathologies and fabricated diagnoses. However, existing medical vision-language models (MedVLMs) are trained predominantly using standard token-level cross-entropy loss, which measures syntactic agreement with ground-truth reports but fails to prioritize diagnostically critical clinical terms. In Diagnostic Loss: Asymmetric Supervision for Radiology Report Generation (arXiv:2610.30980, 22 Oct 2026), researchers Ignacy Kolton, Wenxuan Li, Zongwei Zhou, Jarosław B. Ćwikła, Arkadiusz Sitek, Pedro R. A. S. Bassi, and Szymon Płotka from Jagiellonian University, Johns Hopkins University, GAMMED, and Harvard Medical School introduce Diagnostic Loss (DLoss). DLoss is a differentiable, sequence-level auxiliary training objective that directly optimizes the clinical faithfulness of radiology decoders.

DLoss organizes diagnostic concepts into structured target sets across 11 abdominal CT organ groups (such as liver, gallbladder, kidneys, and pancreas). During training, DLoss applies asymmetric pooling over decoder token probabilities: for findings supported by the ground truth report, a noisy-OR pooling operator requires the model to allocate sufficient probability to the diagnostic concept at at least one token position. Conversely, for unsupported diagnostic alternatives, a log-sum-exp pooling operator penalizes probability spikes across every decoding position. Evaluated across 50,000 abdominal CT scans and multiple 3D MedVLM architectures, DLoss substantially improves diagnostic precision and recall on both in-distribution and out-of-distribution evaluation benchmarks while matching baseline linguistic fluency. Importantly, DLoss modifies only the training loss, requiring no changes to vision backbones, decoders, or test-time inference pipelines.

Why it matters: In radiology, reporting errors carry asymmetric risks: falsely stating that an organ is unremarkable when stones or lesions are present delays treatment, while asserting absent pathologies prompts invasive, unnecessary interventions. Standard language modeling objectives treat every token equally, encouraging models to output fluent, generic prose that often glosses over subtle abnormal findings. By explicitly enforcing asymmetric constraints on clinical concepts, DLoss aligns model optimization with actual diagnostic requirements without imposing runtime latency penalties.

Caveats: The reported evaluation focuses on 3D abdominal CT scans across 11 predefined organ categories and requires pre-extracted diagnostic vocabularies during training. Extending DLoss to rare incidental findings, multi-organ systemic pathologies, or unconstrained free-text reporting across diverse imaging modalities like thoracic MRI or radiographs will require broader diagnostic ontology mapping.