Patients increasingly read their own CT reports. Those reports are written for physicians, and current simplifiers handle them in one of two ways. Some shorten and swap vocabulary, which strips out the information a layperson needs to judge how serious a finding is. Others add explanatory background, which means inventing facts. This note proposes a third option. When a simplifier needs to add context for a less-literate reader, it may add only facts measured from the 3D scan itself: how large a lesion is, how many centimeters it spans along the body, which neighboring organs it touches, and how an organ's size compares to normal. A 3D segmentation model turns the scan into a short "fact sheet." A simplifier whose reading level can be dialed up or down then rewrites the report. At simpler levels it is allowed, and trained, to elaborate, but only with fact-sheet entries. A checker rejects any added claim the fact sheet does not support.
The motivation is a tension in the simplification literature. Elaborative Simplification for German-Language Texts found that comprehension often requires adding background and causal links, and that factuality drops sharply when language models are made to do so. Two-Pronged Human Evaluation of ChatGPT Self-Correction in Radiology Report Simplification found that even a careful multi-agent, text-only simplifier lifted layperson severity judgments only from 38.4% to 52.3%. We claim that both failures share one cause. The text contains too little physical, spatial information to convey severity safely, and the missing information is sitting in the volume. The two sentences to tell a colleague are these: simplify radiology reports with a reading-level knob, but let every added explanation be backed by a measurement from the 3D scan. That should raise lay understanding of severity without the hallucinations that unconstrained elaboration causes.
Core Claim
Grounding elaborations in 3D-derived measurements, rather than in the language model's prior knowledge, should do three things at matched reading level. It should improve lay severity comprehension, reduce unsupported claims, and preserve clinically required negative findings. The benefit should specifically depend on volumetric information, so it should shrink when the same measurements are estimated from a single 2D key slice.
The claim is deliberately narrow about architecture. We do not claim that an end-to-end 3D multimodal language model is needed. Exploring the Design Space of 3D MLLMs for CT Report Generation reports that scaling the language model from 2B to 14B parameters did not improve 3D report quality and that frozen decoders work well. That suggests the bottleneck is what visual evidence reaches the text generator, not how the generator and the image encoder are fused. We therefore pass the volume to the text model as structured, checkable measurements.
Relation to Prior Work
The closest system is ReXplain, which simplifies chest CT reports with GPT-4o, maps findings to 201 anatomical regions, segments those organs with a text-prompted 3D segmenter, and renders a narrated video. ReXplain's simplification never sees the volume. The scan is only used downstream, to draw the organ the text already named. ReXplain also produces one fixed reading level and reports 85% abnormality recall in the text step. We reverse the direction of information flow, so that the volume supplies content to the text, and we add graded control over reading level.
Text-only radiology simplifiers work on isolated sentences without images:
- Data Augmentation for Radiology Report Simplification combines GPT-3 augmentation, a BERT entailment filter and ontology substitution, and reaches SARI 0.5875 on 540 liver CT sentences.
- The multi-agent self-correction study above uses 40 MIMIC-III sentences.
Both motivate the proposal: they show that text-only simplification plateaus. We reuse their entailment-filter idea, but check against the fact sheet instead of the source sentence.
On the imaging side, three systems generate specialist-level text only:
- AutoRG-Brain prompts a report generator with 3D segmentation masks.
- CT-CLIP aligns whole chest volumes with reports.
- The 3D multimodal design study above generates reports on AMOS-MM.
AutoRG-Brain supplies the method pattern of segment-then-describe, which we borrow. The design study also constrains us: it found that appending normal-organ statements raised the clinical GREEN score from 0.366 to 0.470. Completeness therefore matters and must survive simplification.
Control over reading level comes from two papers:
- Controllable Sentence Simplification (ACCESS) uses ratio control tokens for length, paraphrase, word rarity and syntactic depth, and reports SARI 41.87 on WikiLarge.
- A Framework for Fine-Grained Complexity Control in Health Answer Generation fits a 21-level health-text complexity score from 166 features and reports control codes beating few-shot prompting by 23% in mean absolute error.
We reuse both. We add one new control, an elaboration budget, and one hard constraint, a completeness floor.
Two further papers motivate design choices. Your other Left! shows that 2D vision-language models are near chance on relative position in axial CT and fall back on textbook priors. This is why we do not let a vision-language model describe spatial relations from pixels. Deep convolutional networks for pancreas segmentation in CT imaging shows why slice-wise processing distorts extent.
Research Program
Fact sheet. The first component turns a CT volume plus its report into a fact sheet. It works as follows:
- An organ segmenter produces 3D masks. Candidates are the 3D U-Net that reached 89.04% mean Dice on AMOS, or TotalSegmentator-style masks for chest CT-RATE.
- For each finding the report mentions, a text-prompted lesion localizer, seeded by the report's location phrase, produces a lesion mask.
- From these masks we compute a small fixed vocabulary of measurements: lesion diameter and volume, craniocaudal extent (the head-to-foot length along the body axis, in centimeters and in slices spanned), contact with named neighbor structures, organ volume relative to a reference distribution, and bilaterality.
- Every entry carries a confidence derived from mask stability under test-time augmentation. Low-confidence entries are withheld.
Simplifier. The second component is an instruction-tuned open model of about 8B parameters (Llama-3.1-8B-Instruct, as in the health-control paper). It takes the report, the fact sheet and control codes. The codes are:
- a target complexity level on the 21-level scale;
- ACCESS-style ratios;
- an elaboration budget, meaning the number of added explanatory clauses;
- optionally, a personal vocabulary profile built with the active-learning complex-word identification of Gooding et al., which needs about 23 annotations per user.
Training data comes from three sources. We create lay rewrites of CT-RATE and AMOS-MM findings at three target levels. A strong model drafts them, radiologists correct them, and every added clause is tagged with the fact-sheet entry it uses. We augment with the multi-level health corpora from the health-control paper for general control behavior.
Checker. At decoding time, a checker built like the BERT-Checker entailment model, but trained to test whether each added clause is supported by fact-sheet entries, rejects and resamples unsupported elaborations. The completeness floor requires that report statements tagged as clinically required negatives survive in some simplified form. Examples are "no aortic dissection" and "no enlarged lymph nodes."
Main quantity. Our central factuality measure is the grounded elaboration rate. For an output $y$ containing added clauses $E(y)$,
$$\mathrm{GER}(y) = \frac{|\{e \in E(y) : e \text{ is judged supported by the volume}\}|}{|E(y)|},$$
where "supported by the volume" is judged by a radiologist looking at the scan, not by our own checker. This prevents the method from grading itself.
Baselines
Report-only reference. The do-nothing reference is the original radiologist report shown to lay readers. It anchors the comprehension scale; the self-correction study measured 38.4% severity accuracy on MIMIC-III sentences for original text.
Text-only self-correction simplifier. This is the Generator/Radiologist/Patient/Processor chain-of-thought self-correction method from Two-Pronged Human Evaluation. It is the strongest text-only competitor on exactly our outcome, lay severity comprehension, with 52.3% reported. That number comes from different sentences and a different participant pool. We will rerun it on our CT-RATE and AMOS-MM test findings, with the same underlying model family and a prompt-tuning budget equal to our control-code tuning. Beating it means higher severity accuracy at statistically indistinguishable complexity scores and radiologist correctness of at least 4.5 out of 5.
Text-only control codes. This is the fine-tuned control-code model from A Framework for Fine-Grained Complexity Control, trained on our same lay rewrites but with the fact sheet removed and elaborations allowed. The paper reports only relative gains in mean absolute error: 23% below few-shot, and nearly 50% below baseline. We will treat our reproduced value on our data as the bar our method must match on reading-level control while beating it on GER.
ReXplain-style pipeline. This is GPT-4o simplification from text, as in ReXplain, which reports 85% abnormality recall. We run it on the same CT-RATE cases and score recall, GER and comprehension.
Ablations. Three ablations isolate the key ingredient:
- The same fact sheet computed from the single 2D key slice with the largest lesion cross-section. This isolates the volumetric contribution.
- The fact sheet with shuffled or perturbed values. This tests whether the model actually reads the sheet.
- Removal of the checker. This separates training-time grounding from decoding-time filtering.
Why not just…?
Why not just generate a clinical report with a 3D model and have an LLM simplify it? When the radiologist's report exists, a generated report adds nothing. The question is whether the simplifier has anything beyond the words. The text-only control-code baseline is exactly this alternative, given the true report rather than a generated one. It gives the alternative its best case. If it matches our method on severity comprehension and GER, the proposal's claim fails, and we will say so.
Why not just show a 2D VLM the key slice and ask for a simple explanation? Your other Left! reports near-50% relative-position accuracy for GPT-4o and similar models on axial CT, falling from 75% to 50% under flipping. That makes free-form spatial explanation from pixels unsafe. Still, a key slice might be enough for size, so we test the fairer version: key-slice-derived measurements in the same fact-sheet format. If the 2D-sheet ablation matches the 3D sheet, volumetric context adds nothing to lay explanation, and the simpler system should be preferred.
Why not just paste measurements into a GPT-4o prompt? This is a real threat, because it is close to our method without the training or the checker. We include it as a baseline: frontier model, fact sheet in the prompt, reading level requested in words. The health-control paper's advantage for control codes over prompting predicts looser level calibration. Whether prompting also has a lower GER is an open empirical question that the experiment settles. If prompted GPT-4o with the sheet matches us, the contribution reduces to the fact sheet itself, which is still a finding worth reporting.
Proposed Experiments
Data. Test sets are held-out CT-RATE chest cases and AMOS-MM abdominal cases, about 300 findings each, balanced across normal, incidental and serious findings.
Measures. Lay participants recruited at three screened literacy levels read one system's output per case and answer four things:
- severity category;
- body location;
- whether follow-up is needed;
- a multiple-choice question about extent.
Radiologists who can view the volume rate correctness and completeness, and label each added clause for GER. Automatic measures are:
- reading-level mean absolute error under the 21-level scorer;
- SARI against the corrected lay references, reported but not optimized, given the known mismatch between overlap metrics and correctness;
- retention of required negatives;
- abnormality recall against the report.
Decisive checks. There are three:
- Severity and extent accuracy must beat the self-correction and text-only control-code baselines at the lowest literacy tier, while GER is higher than for text-only elaboration.
- The 3D sheet must beat the 2D-sheet ablation specifically on extent and multi-organ findings. If it does not, the volumetric claim is refuted.
- Perturbing the fact sheet must change the output accordingly. If it does not, the model is ignoring the scan.
Robustness. We will inject segmentation errors at controlled rates to measure how quickly lay understanding degrades. This sets the segmentation accuracy the approach requires before it could be used with patients.
References
- Elaborative Simplification for German-Language Texts
- Two-Pronged Human Evaluation of ChatGPT Self-Correction in Radiology Report Simplification
- Exploring the Design Space of 3D MLLMs for CT Report Generation
- ReXplain: Translating Radiology into Patient-Friendly Video Reports
- Data Augmentation for Radiology Report Simplification
- AutoRG-Brain: Grounded Report Generation for Brain MRI
- A foundation model utilizing chest CT volumes and radiology reports for supervised-level zero-shot detection of abnormalities
- Controllable Sentence Simplification
- A Framework for Fine-Grained Complexity Control in Health Answer Generation
- Your other Left! Vision-Language Models Fail to Identify Relative Positions in Medical Images
- Deep convolutional networks for pancreas segmentation in CT imaging
- AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation
- One Size Does Not Fit All: The Case for Personalised Word Complexity Models