Lung Nodule Lung-RADS Risk Classifier

From LIDC-IDRI to Lung-RADS: Two Answers to the Same Problem

Long before Lung-RADS existed as a standardized reporting tool, researchers building computer-aided nodule detection needed something more basic: a large, carefully labeled dataset of real CT scans showing exactly where nodules were and how confident different experts were about them. That need produced LIDC-IDRI, the Lung Image Database Consortium and Image Database Resource Initiative, a public collection of thoracic CT scans in which pulmonary nodules were independently annotated by up to four experienced thoracic radiologists per case, without those radiologists seeing each other's markings first. That "independent, blinded, multi-reader" design was deliberate, and it revealed something uncomfortable: even trained radiologists frequently disagree, sometimes substantially, about a nodule's precise boundary, whether a borderline finding even counts as a true nodule at all, and how likely it is to be malignant. This inter-reader variability is not a flaw in any individual radiologist's skill, it is a genuine reflection of how ambiguous small pulmonary findings on CT can be. LIDC-IDRI captured that variability explicitly rather than papering over it, which made the dataset unusually valuable for training and validating machine-learning models meant to detect nodules and estimate malignancy risk, since a model tested against a single reader's opinion risks simply learning that one radiologist's idiosyncrasies rather than a generalizable signal. Lung-RADS tackles a closely related but distinct problem from the clinical side. Rather than building better detection algorithms, the American College of Radiology's Lung-RADS system standardizes how any radiologist, anywhere, reports a nodule they have already found, turning what used to be free-text impressions and inconsistent follow-up recommendations into a small, discrete set of categories, each tied to a specific, defensible, evidence-based follow-up interval. Where LIDC-IDRI documents and quantifies human disagreement to build better AI, Lung-RADS reduces the practical consequences of that same disagreement in day-to-day screening care. This simulation focuses on the Lung-RADS side: given a nodule's characteristics, what category does it fall into, and what does that category actually recommend?

The Lung-RADS Categories: Size, Composition and a Simple Escalation Ladder

Lung-RADS v1.1 organizes nodules into a small number of assessment categories, and this simulator focuses on the four that matter once a nodule has actually been found: Category 2, 3, 4A and 4B. Category 1, a fully negative scan with no nodules at all, is not something a nodule-diameter slider can represent, so it is not modeled here. Category 2 covers nodules with a real, published, very low likelihood of becoming a clinically significant cancer, based specifically on their small size or stable behaviour over time. For a solid nodule, that means a diameter under 6mm at baseline. The recommended action is simply to continue with annual screening, a 12-month low-dose CT, the same interval as if nothing concerning had been found at all. Category 3 flags a probably-benign finding: for a solid nodule, that is a diameter from 6mm up to just under 8mm. Rather than waiting a full year, Lung-RADS recommends a 6-month low-dose CT to see whether the nodule changes, since a modest size at this range still carries enough uncertainty to warrant closer monitoring than the standard annual interval, without yet justifying invasive follow-up. Category 4A is where Lung-RADS starts recommending genuinely more urgent action: solid nodules from 8mm up to just under 15mm fall here, and the recommendation becomes a 3-month low-dose CT, with PET/CT sometimes considered when the solid component reaches 8mm or more. This is a meaningful escalation from a passive "wait and re-scan in six months" to active short-interval monitoring. Category 4B is the most suspicious category modeled here: solid nodules at 15mm or larger. The published recommendation shifts from simple repeat CT to diagnostic chest CT with or without contrast, PET/CT, and/or tissue sampling such as a needle biopsy, reflecting that a nodule this large carries a substantially higher real probability of malignancy and needs a more definitive answer than another wait-and-watch scan can provide. Set the diameter slider in the simulation across this range with "Solid" selected and watch the category, its color, and its recommended interval change at exactly these published cut-points, 6mm, 8mm and 15mm.

Why Composition Changes Everything: Solid vs. Part-Solid vs. Ground-Glass

Not every nodule looks the same on CT, and Lung-RADS explicitly accounts for that by applying different real risk logic to different nodule densities, because a solid nodule and a hazy ground-glass nodule of identical diameter carry meaningfully different malignancy probabilities in practice. A solid nodule is one that completely obscures the underlying lung tissue on CT, judged on its full diameter using the 6/8/15mm cut-points described above. A part-solid nodule has both a solid component and a hazy ground-glass component. Because the solid portion correlates more strongly with invasive disease than the ground-glass portion does, the full published Lung-RADS criteria assess part-solid nodules primarily on the size of that solid component rather than the nodule's overall diameter. This simulation, for simplicity, uses the diameter slider as a stand-in for that solid-component size when "Part-solid" is selected, which is a deliberate simplification worth being explicit about: real part-solid assessment separately measures the solid portion inside the larger part-solid nodule, something a single diameter control cannot fully capture. A pure ground-glass nodule (also called nonsolid) has no solid component at all, appearing only as a hazy density on CT. These nodules behave differently enough, and typically far more indolently, that the published Lung-RADS criteria cap them at Category 2 below 30mm and Category 3 at or above 30mm, without escalating to 4A or 4B on size alone, reflecting their comparatively slower, less aggressive typical clinical course even when relatively large. Switch the composition selector in the simulation between solid, part-solid and ground-glass at the same diameter and watch how differently each is categorized, a direct illustration of why "how big is it" is never a complete question in nodule assessment.

Growth: When Change Over Time Outweighs Absolute Size

A nodule that has not changed size across two scans separated by months is reassuring in a way that raw diameter alone does not capture, and one that has visibly grown is concerning in a way that raw diameter alone does not capture either. Lung-RADS formalizes this with a precise, published definition: growth means an increase in nodule diameter of at least 1.5 millimetres compared with the prior exam. A nodule that meets this growth definition gets its reported category upgraded, independent of what its size-based category would otherwise have been. A growing nodule that is still under 8mm becomes Category 4A, triggering the 3-month follow-up interval regardless of how reassuring its absolute size might have looked in isolation. A growing nodule that has reached 8mm or larger becomes Category 4B, triggering the most urgent recommended workup. This growth-based override exists because a demonstrated trajectory, a nodule actively getting bigger between two real scans, is itself independent evidence, arguably stronger evidence in some respects than a single static size measurement taken at one point in time. Toggle the follow-up behaviour selector to "Growing" in the simulation and adjust the growth-rate slider around the 1.5mm/year threshold to see this upgrade rule apply directly, including the visual dashed ring the simulation draws around a growing nodule and the note confirming whether the simulated growth meets the real published Lung-RADS growth definition.

Why Standardization Like This Matters for Screening Programs

Before structured systems like Lung-RADS became widespread, nodule follow-up recommendations varied considerably between radiologists and institutions even for very similar findings, driven by exactly the kind of inter-reader variability that datasets like LIDC-IDRI were built to document. That inconsistency has real downstream costs: patients with genuinely low-risk nodules sometimes received unnecessarily frequent, anxiety-inducing and expensive follow-up imaging or even invasive biopsies, while some higher-risk findings received less urgent follow-up than they warranted. By tying every nodule's follow-up recommendation to a small set of explicit, published, size- and behaviour-based rules, Lung-RADS reduces that variability directly, giving patients across different hospitals and different reading radiologists a more consistent, evidence-anchored experience regardless of who happens to interpret their scan. It also gives screening programs a structured audit tool: because every scan gets one of a small number of discrete categories, programs can track outcomes by category over time and validate, in aggregate, whether the recommended intervals are actually striking the right balance between catching real cancers early and avoiding unnecessary imaging for the many nodules that are never going to become clinically significant. This simulation is a compact way to explore that same underlying logic interactively, seeing exactly which combination of size, composition and growth pushes a simulated nodule from a reassuring annual-screening category into one recommending closer, more urgent follow-up, and why each of those published thresholds sits where it does.

Frequently asked questions

Try it live

Everything above runs in your browser — open Lung Nodule Lung-RADS Risk Classifier and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab. Remember: this is an educational tool illustrating a published classification framework, not a diagnostic instrument.