Lung-RADS: Standardized Lung CT Screening

• Lung-RADS is a standardized framework that categorizes lung nodule features from CT scans to guide risk assessment and follow-up recommendations.
• Integration of multimodal data and advanced AI techniques enables continuous, individualized risk scoring with enhanced predictive accuracy.
• Clinical translation of Lung-RADS supports personalized management decisions, though further research is needed to overcome limitations in data scope and sensitivity.

The Lung CT Screening Reporting and Data System (Lung-RADS) is a standardized framework designed to stratify and manage risk in lung cancer screening using computed tomography (CT). It provides a categorical system for nodule assessment based primarily on radiological features, guiding patient management and follow-up recommendations. The recent literature reveals Lung-RADS as both a clinical anchor and a reference standard for AI-augmented risk estimation, yet also points to areas where its limitations—chiefly the exclusive reliance on imaging findings—may be mitigated by integrative, data-driven approaches.

1. Principles and Scope of Lung-RADS

Lung-RADS formalizes the reporting and management of findings in lung cancer screening CT studies by assigning categories based on a constellation of radiological descriptors including nodule size, attenuation, margin, location, and interval growth. The categorical scale (e.g., Lung-RADS 0, 1, 2, 3, 4A, 4B, 4X) is determined according to prespecified criteria. These categories drive recommendations for surveillance interval, additional imaging, or biopsy. As a qualitative tool, Lung-RADS establishes a common language for both clinical and research settings, supporting outcome audits and risk communication.

Recent works have emphasized the need to improve sensitivity and specificity of Lung-RADS, noting that rigid categorization can lead to suboptimal risk stratification, particularly in patients with atypical or borderline findings (Niu et al., 7 Sep 2025).

2. Integration of Multi-modal Data and Advanced AI Methods

The latest approaches extend the Lung-RADS paradigm by integrating additional modalities such as clinical history, genomics, and laboratory markers into the risk assessment, using LLMs that reason over diverse data sources (Niu et al., 7 Sep 2025).

A "reasoning LLM" (RLM) accepts structured imaging descriptors along with longitudinal risk factors—demographics, smoking history, family history, occupational exposures—as input. Text templates and augmentation are used to ensure consistency across multimodal data, and supervised fine-tuning with distillation and reinforcement learning enables explicit chain-of-thought reasoning. The model decomposes the risk task, analyzing each factor, then synthesizes these into a final score via a data-driven system equation:

$$P_\theta(r, y \mid x) = P_\theta(r \mid x) \cdot P_\theta(y \mid x, r)$$

where $r$ corresponds to the reasoning trace and $y$ to the answer (risk score $s$), supporting monitorability and clinical interpretability. This augmentation directly addresses Lung-RADS limitations by allowing continuous, individualized risk scoring rather than rigid categorization.

3. Risk Stratification and Predictive Performance

Whereas standard Lung-RADS uses static thresholds for nodule size and characteristics, advanced AI models dynamically weight the importance of diverse risk contributors, optimizing sensitivity and specificity as quantified by the area under the receiver operating characteristic curve (AUC).

The RLM, using chain-of-thought reasoning and explicit sub-task decomposition, achieves superior short-term prediction AUC of up to 0.926 for one-year risk, outperforming conventional Lung-RADS categorization and general-purpose baseline models (AUC 0.54–0.61) across all prediction horizons (Niu et al., 7 Sep 2025). The reward functions for reinforcement learning are explicitly engineered to calibrate output scores and format, using piecewise functions and length penalties to enforce correct prediction boundaries and report structure.

Score reward function:

For $\ell = 0$ (negative ground truth):

• if $s \leq t_1$, reward = 1
• if $s > t_1$, reward = $1 - 2s$

For $\ell = 1$ (positive ground truth):

• if $s \leq t_2$, reward = $2s - 1$
• if $s > t_2$, reward = 1

with $t_1 = 0.45$, $t_2 = 0.55$.

Final reward:

$$f_\text{reward} = \alpha f_\text{score} + \beta f_\text{format} + f_\text{length}$$

where $\alpha = \beta = 1$.

Monitorability is enhanced by outputting explicit reasoning pathways (> tokens) for every prediction, allowing clinicians to verify which factors drive the risk score.

4. Data Inputs, Reasoning Pathways, and Case Example

The RLM framework formalizes the integration of imaging and non-imaging predictors through tokenized free-text templates, processed and parsed by the model. For instance, an LDCT report describing a "new spiculated 6 mm nodule" plus "heavy smoking history" and "interstitial changes" is decomposed into its components.

The chain-of-thought output would sequentially assess malignancy risk for the new nodule (considering margin and size), adjust for smoking intensity and duration, and evaluate the presence of interstitial findings. Each sub-component's contribution to the total risk score is transparent, supporting targeted review by clinicians. Final output is presented in the format $s = g(y)$, with deterministic extraction of the numeric score from the answer tokens.

A case example cited in the paper demonstrates this decomposition and reasoning trace, with the final risk presented as a boxed LaTeX number.

5. Clinical Translation and Implications

The transparent and individualized scoring system supports new modes of patient management, enabling nuanced decisions such as the timing of surveillance imaging, need for PET-CT, or biopsy referral. Instead of binary thresholds (e.g., Lung-RADS 3 versus 4A), the continuous risk score supports personalized action aligned with precision medicine principles.

Interpretability—through chain-of-thought reasoning—fosters trust in the underlying AI and opens avenues for iterative improvement when coupled with feedback from domain experts.

The curated dataset from the National Lung Screening Trial (NLST) ensures real-world relevance and robust validation. By mapping risk scores to actionable outcomes, and allowing dynamic updates as new longitudinal data accrue, the RLM approach moves Lung-RADS towards a more adaptive, evidence-driven screening tool.

6. Limitations and Future Directions

The data-driven RLM approach depends on the breadth and fidelity of input data; limited or biased datasets may restrict generalizability. Reward hacking near boundaries and the possibility of mis-weighted factors, though partially mitigated by the designed reward functions, require ongoing surveillance and clinical review.

Further research aims at extending the RLM to support real-time integration with electronic medical records and streamlining deployment within PACS/RIS environments, subject to regulatory validation and cross-institutional harmonization.

A plausible implication is that chain-of-thought enabled risk assessment—coupled to the original Lung-RADS—may become the reference standard for future high-throughput lung cancer screening programs, integrating imaging and non-imaging data and supporting iterative, transparent decision-making.

---

In summary, Lung-RADS, originally a categorical imaging-based system, is positioned as a clinical and research anchor for lung cancer screening. Its future is increasingly linked to integrative, reasoning-enabled models such as the RLM, which offer more granular and transparent risk assessment by incorporating diverse clinical and imaging predictors using a rigorously engineered data-driven system equation and reward structure. Performance improvements, monitorability, and case deconstruction illustrate the practical impact of this evolution. (Niu et al., 7 Sep 2025)