Generating Reports or Repeating Templates? Measuring and Mitigating Template Collapse in 3D CT Report Generation

Researchers identify 'Template Collapse' as a critical failure mode in 3D medical imaging AI systems, where vision-language models generate fluent but clinically inaccurate reports that miss rare pathologies. They propose CLarGen, a decoupled framework that separates pathology detection from language generation, achieving significant improvements in clinical accuracy metrics while maintaining report quality.

The paper addresses a fundamental problem in medical AI deployment: systems that appear to work well on surface metrics fail catastrophically in clinical practice. Modern 3D CT report generation models achieve fluent text output but systematically under-detect critical findings, particularly rare conditions that carry high clinical stakes. This disconnect between linguistic quality and medical accuracy stems from how these systems are trained—text generation objectives reward plausible-sounding templates rather than precise clinical grounding.

The root causes are endemic to medical AI development: limited training data, severe class imbalance favoring common diagnoses, and weak signal extraction from volumetric imaging data. Under these constraints, models learn shortcuts that sacrifice accuracy for fluency. The research systematically measures this failure through clinical fidelity scores, output diversity metrics, and rare-finding detection rates—establishing benchmarks for evaluating what was previously unmeasured.

CLarGen's decoupled approach represents a significant methodological shift. By explicitly separating what-to-detect (pathology identification via Latent Query Transformer) from how-to-communicate (language synthesis from detected findings), the framework enforces clinical grounding as a hard constraint rather than a soft objective. Results demonstrate macro-F1 improvements from 0.189 to 0.487 and clinical report generation scores from 0.368 to 0.472, substantial gains in medical accuracy.

For the medical AI industry, this work highlights why end-to-end models trained on language objectives alone are insufficient for safety-critical applications. The framework's modular design enables better interpretability and clinical validation. As medical institutions increasingly deploy AI diagnostic assistants, this research suggests that explicit clinical reasoning components—not just language sophistication—are prerequisite for trustworthy systems.

• →Template Collapse reveals fundamental misalignment between linguistic fluency and clinical accuracy in medical AI systems.
• →CLarGen's decoupled architecture separates pathology detection from report generation, achieving 2.5x improvement in macro-F1 scores.
• →Medical AI training objectives must enforce explicit clinical grounding rather than relying on end-to-end language modeling.
• →Limited data and label imbalance in medical imaging create conditions where models learn to prioritize common findings over rare critical pathologies.
• →Systematic measurement of clinical fidelity, diversity, and rare-finding survival enables previously unmeasurable quality assessment in medical reports.