Reading the Report as a Checklist: Text-Triggered High-Resolution Re-Inspection for Small CT Lesions

Lesion detectors for CT scans usually work in two steps. A first network proposes candidate boxes where something lesion-like might be, and a second network looks closely at each candidate to classify it and tighten the box. Small lesions break this design in a predictable way. Before the proposal step ever sees the image, the backbone shrinks it repeatedly, and a lesion a few millimetres wide can be blurred into the background. If no candidate box is proposed, the careful second step never runs. Yet for most clinical scans a human has already written down what is there. The radiology report says, in free text, something like "4 mm hypodensity in segment VII of the liver," and for follow-up scans the prior report lists the lesions being tracked.

We propose to use that text as a checklist rather than as a prompt. An extractor turns the report into an inventory of findings, each with an organ, an approximate size and a count. The inventory is reconciled against the detector's first-stage proposals. Any finding the detector failed to account for triggers a second, local search. That search runs at full native resolution inside the named organ, with the backbone's downsampling reduced and with candidates filtered by the reported size. The claim is narrow and testable. Report text is most valuable for deciding where to spend expensive high-resolution computation, and the gains should be concentrated on exactly the lesions that downsampling erases: small ones, at a fixed false-positive budget.

What exists and where this differs

Text and detection have been combined before, but never in this direction. SP-Det generates a synthetic report from a chest radiograph with a vision-language model and feeds parsed text prompts into a cross-modal detector, reaching 21.0% mAP40:95 on VinDr-CXR. Its text is invented by the model itself, and it does nothing about resolution. MedRG grounds phrases from real reports onto chest X-rays with a box token and a high-resolution SAM encoder. It works on 2D radiographs and has no proposal-refinement stage. Cheng et al.'s unified DeepLesion framework runs detection, then crop-based segmentation, then report generation, so text appears only as an output. The authors' own pipeline fails completely when the first-stage detector misses a lesion. These three papers motivate the proposal: text helps detection, and first-stage misses cannot be recovered.

Several papers supply the method. The DeepLesion detectors (Yan et al., JMI; Yan et al., arXiv) show that removing one pooling layer to keep feature maps larger improves accuracy by roughly 15%. Cheng et al. use space-to-depth convolutions, which downsample without discarding pixels. Both are evidence that resolution is the limiting factor for small lesions. Our re-inspection head reuses these modifications, but only inside triggered regions, where their cost is affordable. Resolution Meets Reduction shows that organ-guided cropping improves clinical F1 in 19 of 20 report-generation configurations. It also reports a "resolution ceiling": increasing input resolution stops helping once it exceeds the encoder's native scale. That second result is a constraint on our design. Re-inspection therefore works at native resolution with reduced network stride, and does not simply upsample the crop.

On the text side, mCODEGPT offers two lessons. Its two-phase parser first asks whether each entity is present and only then extracts values; with GPT-4o this reaches 93.9% accuracy with a 1.1% hallucination rate on synthetic oncology notes. Its top-down pruner, by contrast, skips every child entity when it misses a parent term. We adopt the two-phase design and avoid top-down pruning. CheXbert and Yan et al.'s RadLex mining supply a trained-extractor alternative and the lesion–sentence pairs we need. The RadLex work also warns that reports omit findings. That warning is why the inventory can only add search effort and never suppress a proposal.

Core Claim

Given a detector with a fixed false-positive rate per image, triggering high-resolution re-inspection from report-extracted findings will raise sensitivity on lesions under about 10 mm. The gain should be substantially larger than what the same extra computation buys when it is spent without text, for example on uniform tiling or on every organ crop. Lesions above roughly 20 mm should show almost no gain, because downsampling does not erase them. If the gain is flat across lesion sizes, or vanishes once compute is matched, the claim is wrong.

Research Program

The system has four parts.

Stage-1 detector. This is a standard first-stage proposal network built on 3DCE-style multi-slice context (3DCE).

Finding extractor. This reads report sentences and outputs a list of (organ, size, count, attributes) entries using two-phase querying. We compare an LLM version against a CheXbert-style fine-tuned extractor trained on the RadLex-mined sentences.

Reconciler. This assigns each first-stage proposal to an organ using a TotalSegmentator organ mask, as Cheng et al. and Resolution Meets Reduction both do. It then matches proposals to inventory entries by organ and by size compatibility.

Re-inspection head. For each unmatched entry, this head runs proposal generation on the native-resolution organ crop. The head has the late pooling removed and uses space-to-depth downsampling. Its anchors are centred on the reported size, and its confidence threshold is lowered. Its candidates then pass through the same second-stage classifier as ordinary proposals, so the precision of the second stage still applies.

We study two text settings. In the same-study setting, the report for the scan being analysed guides detection. This is a legitimate task for retrospective dataset curation and for checking agreement between image and report. Its risk is optimism, because DeepLesion bookmarks were drawn while those reports were written. In the prior-study setting, the report from the patient's earlier exam guides detection on the follow-up scan. DeepLesion's 10,825 studies from 4,477 patients make this possible. It is the deployable case, and it is our headline. A third analysis uses the "all-lesion" re-annotated test set from 3DCE. Its purpose is to check that text guidance does not lower sensitivity to lesions the report never mentions.

Baselines

The naive reference is the Stage-1 detector run at ordinary resolution with no text and no re-inspection. For context, 3DCE reports 84.37% sensitivity at 4 false positives per image with a 27-slice context, against 80.32% for a 3-slice Faster R-CNN, and 82.75% on the all-lesion re-annotated test set. We will reproduce our own backbone's number rather than rely on these. We will report results across the whole range of false-positive rates, because the claimed gains should appear at strict operating points such as 0.5 and 1 false positive per image, where the published 4-false-positive figure hides them.

The strongest existing resolution-preserving detector is YOLO-TLP-MOE from Cheng et al., which reports an mAP50 of 70.1% on DeepLesion. We also include the pool4-removed Faster R-CNN from Yan et al., which reports 64.30% accuracy at IoU 0.5. Both will be retrained on our exact split and evaluated with our sensitivity metric, using the same augmentation and an equal hyperparameter search budget. Beating them means higher sensitivity on sub-10 mm lesions at 1 false positive per image without losing sensitivity on large lesions.

The strongest text-conditioned competitor is an SP-Det-style detector (SP-Det). It receives an embedding of the same report as a global prompt fused into the backbone, but does no reconciliation and no re-inspection. It is given the same text and the same training data. No result is reported on Lacuna for SP-Det or for MedRG on DeepLesion CT. For reference, SP-Det reports 21.0% mAP40:95 on VinDr-CXR and MedRG reports 47.65% mIoU on MRG-MS-CXR.

Three ablations isolate the key ingredient. Shuffled reports use another patient's inventory, which keeps the amount of extra computation identical while removing the information. Organ-only drops the size prior. Oracle inventory uses manually verified findings, which bounds how much extraction errors cost.

Why not just…?

Why not just run the high-resolution detector everywhere? This is the most serious threat. Full-resolution tiling with reduced stride might recover the same small lesions with no text at all. The report contributes no pixels. What it contributes is a restriction on where to look and what size to expect, which should pay off in false positives rather than in raw recall. We will run uniform high-resolution tiling at equal compute and also at unlimited compute, and compare sensitivity at fixed false-positive rates. If unlimited tiling matches our sensitivity at the same false-positive rate, the text adds only efficiency, and we will report it that way.

Why not just crop every organ at high resolution? Resolution Meets Reduction shows that organ cropping alone helps a great deal, so a crop-all-organs re-inspection is a strong text-free control. Our prediction is that it buys recall at a higher false-positive cost, because most organs contain no lesion. The shuffled-report ablation and the crop-all control together separate the benefit of the extra resolution from the benefit of the report's information.

Why not just feed the report as a prompt to an end-to-end detector or grounding model? SP-Det and MedRG show that prompting helps on radiographs. A global text embedding cannot, however, restore features that the backbone's downsampling has already destroyed, and the 3D report generators that rely on heavy token compression lose small focal findings (Resolution Meets Reduction). The SP-Det-style baseline tests this directly. If it matches us on sub-10 mm lesions, reconciliation is unnecessary.

Why not just use a rule-based extractor? We may be able to. Organ and size mentions are simpler to extract than the hedged findings that cost rule-based labelers 0.743 F1 against CheXbert's 0.798 (CheXbert). We will include a regular-expression extractor. If detection sensitivity is equal across extractors, we will recommend the simple one.

Proposed Experiments

All experiments use DeepLesion with patient-level splits, the ~20,000 lesion–sentence pairs, and the 1,872 verified test lesions from Yan et al. 2019. The primary metric is free-response sensitivity at 0.5, 1, 2 and 4 false positives per image, stratified by long-axis diameter into the bins under 5 mm, 5–10 mm, 10–20 mm and over 20 mm, and also by slice thickness. The thickness split matters because 3DCE reports degradation on thick-slice scans. Secondary metrics are extraction precision and recall for organ and size against verified sentences, and high-resolution pixels processed per study.

The decisive check is the interaction between lesion size and text guidance. The real-report minus shuffled-report sensitivity gap at 1 false positive per image should be largest below 10 mm and near zero above 20 mm. A flat profile would mean we are buying compute, not information. The second check is the prior-study setting. It must retain a meaningful share of the same-study gain, or else the method is only a curation tool. The third check is safety. Sensitivity on the all-lesion set's unmentioned lesions must not drop relative to the naive detector. The fourth check is the oracle-versus-extracted gap, which tells us whether extraction or detection is now the bottleneck.