Where Scanner Shift Hurts Most: Lymph Nodes in Unexpected Places

Detectors that find enlarged or metastatic lymph nodes on CT are judged on nodes in the places cancer usually spreads first. Cancer often breaks that order. Papillary thyroid cancer reaches the lateral neck nodes while the central nodes stay clean in roughly one in eight surgical patients. Scrotal lymphoma can appear only in the mediastinum. Kidney cancer can recur in the bowel without passing through the lungs. These "skip" or atypical lesions are the ones a detector most needs to catch and is least prepared for. Without nearby diseased nodes or a familiar nodal station to lean on, the detector has to recognize the node from its local appearance alone. Local appearance is exactly what changes when a hospital uses a different CT manufacturer, reconstruction filter, or slice thickness.

This note proposes to test one hypothesis and act on it. The hypothesis is that hardware-induced domain shift causes a disproportionate loss of sensitivity on isolated, atypically located nodes. The fix we propose avoids rewriting images. Instead, the image is described with a frozen pretrained feature extractor, and per-scanner differences are removed in that feature space using only unlabeled healthy tissue as anchors. The tissue anchors come from automatic organ, muscle and fat masks. Because the correction is fitted only on normal tissue, a small node cannot be "harmonized away." We also ask a cross-domain transfer question that decides which feature extractor to use: are features from a model pretrained on natural photographs easier to align across scanners than features from a model pretrained on CT from one hospital?

Core Claim

The claim has three parts. First, cross-scanner sensitivity loss is larger for isolated nodes and for nodes in body regions unseen during training than for nodes clustered in canonical stations. Second, a node scorer built on frozen patch features, corrected per scanner by matching tissue-class feature statistics on unlabeled target scans, recovers most of that atypical-node gap. It should do so without the false-negative risk of pixel-level harmonization. Third, whether natural-image features (DINOv2) or in-domain CT lymph node features align better is an empirical question with a clear test. If the in-domain encoder wins, we adopt it, and the method still stands.

Grounding in Prior Work

Several papers motivate the gap. Zhu et al. predict skip metastasis in thyroid cancer with a support vector machine on six clinical and pathological variables. They report 13.3% and 11.6% skip prevalence and a validation AUC of 0.721, but they use no imaging. Case reports document cross-compartment bypass with no imaging AI behind it, for example scrotal lymphoma to mediastinum and isolated renal-cell bowel recurrence. The station-aware chest CT detector shows the opposite design choice. Conditioning on anatomical station raises sensitivity from 65.1% to 71.4% at 2 false positives per patient. That is good for canonical nodes, but by construction it rewards context, and context is what an isolated node lacks. We treat it as a competitor whose strength on typical nodes may become a weakness on atypical ones.

Other work constrains the method. LMSB harmonization of retinal OCT shows that pixel translation methods trade tissue boundaries for contrast matching. Its fix needs dense layered anatomy, which nodes do not have. The domain- and geometry-agnostic V-Net combines an adversarial scanner classifier with a shape prior. It works for the left atrium, but metastases have no canonical shape, and its 64-cubed downsampling would erase small nodes. We take from these papers the lesson that the correction must touch neither image pixels nor lesion shape.

Two papers supply the method. Rastogi et al. align RETFound embeddings across two fundus cameras with Wasserstein-trained tabular models, without retraining the foundation model. We move that idea to dense 3D CT patch features, and we condition the alignment on tissue class instead of aligning whole-image vectors. DINOv2 few-shot medical segmentation shows natural-image ViT features adapt to abdominal organs, with Abd-CT mean Dice of 73.72 versus 63.02 for ALPNet. However, that evaluation covers only large organs on a single source. FM-OSD shows that frozen DINO features beat medical self-supervised models for one-shot landmarks. DGST gives the in-domain alternative: an nnUNetv2 lymph node model pretrained on 36,106 RADCURE nodes and fine-tuned few-shot.

Research Program

We first build a stratified evaluation, since no benchmark currently separates atypical nodes from typical ones under scanner shift. From the public CT lymph node datasets in the file (RADCURE, SegRap2023, LNQ2023, ABD-LN, CRC-LNM, and the lymph node subset of DeepLesion), we define scanner domains by manufacturer and reconstruction kernel from DICOM headers where available. Where headers are unavailable, we fall back to dataset of origin, and we report which grouping was used for each split.

True skip labels need primary-site and pathology information that most public sets lack. We therefore use two operational proxies, stated as proxies:

  • An isolated node is a ground-truth node with no other annotated node within a fixed distance or in its station.
  • An unseen-region node is a node in a body region absent from training, following DGST's head-and-neck-to-mediastinum transfer.

A second phase applies the pipeline to a retrospective thyroid cohort with preoperative neck CT and central and lateral dissection pathology, modeled on the Zhu et al. design. There, skip status is real.

The method has three components:

  1. Features. A frozen encoder produces patch features on 2.5D slabs. The encoder is either DINOv2 with a slice adapter, as in Ayzenberg et al., or the DGST/RADCURE encoder. The CT window is mapped into the network input so Hounsfield-unit contrast is not lost.
  2. Node scorer. A light node scorer and box regressor is trained on source scans. During training we randomly mask the scorer's access to coarse location embeddings, so it cannot rely only on station context.
  3. Per-scanner alignment. For each target scanner, TotalSegmentator masks select patches of liver, spleen, muscle, fat, vessel and bone. We fit an affine map that brings target tissue-class feature distributions onto the source ones:

$$\min_{A_t,\,b_t}\ \sum_{k} W_2^2\!\left(A_t Z^t_k + b_t,\ Z^s_k\right)$$

Here $Z^t_k$ and $Z^s_k$ are the patch features of tissue class $k$ on target scanner $t$ and on the source, $A_t, b_t$ are the per-scanner linear map and offset, and $W_2$ is the 2-Wasserstein distance. We compute $W_2$ in its Gaussian closed form, which makes the fit a small least-squares-like problem. No lesion patch enters the fit, so the map can only correct what normal tissue reveals about the scanner.

The cross-domain question becomes measurable through one quantity: the residual $W_2$ after alignment for each encoder. Lower residual means scanner shift is closer to linear in that feature space.

Baselines

The naive reference is nnU-Net trained on the source domain with its default augmentation and applied unchanged to the target scanner. On ABD-LN, LN-DDPM reports this configuration at DSC 0.5273 and NodeRecall 0.6405 within domain. A matching cross-scanner number was not found on Lacuna. This baseline defines the "do nothing" drop.

The strongest existing method for lymph node CT is DGST fine-tuning from the RADCURE foundation model. It reports 68.44% DSC on SegRap2023 at 10 shots and 63.82% DSC on LNQ2023 at 20 shots, against 62.47% for full fine-tuning (DGST). It is the strongest competitor because it combines in-domain lymph node priors with careful few-shot adaptation. We run it in two arms:

  • Zero target labels. Trained on source only, matching our unsupervised setting.
  • k target labels. k of 10 and 20 target-scanner scans, which is DGST's native regime. Our method also gets the same k labeled scans for its scorer.

Both methods receive identical splits, GPU-hour budgets and hyperparameter searches.

For context-reliance we compare against the station-aware detector, which reports 71.4% sensitivity at 2 false positives per patient on its lung cohort (Yan et al.). We reimplement it on our thoracic splits. The prediction is that it is strongest on clustered nodes and weakest on isolated ones.

For generative data scarcity fixes we compare LN-DDPM augmentation (NodeRecall 0.7486 with 10x synthetic data on ABD-LN). For natural-feature adaptation we compare DINOv2 with LoRA and test-time training as in Ayzenberg et al.

Three ablations isolate the key ingredients:

  • Alignment removed.
  • Alignment fitted on all patches rather than tissue-only.
  • Location masking removed.

We also swap the encoder between DINOv2 and RADCURE.

Beating the strongest baseline means two things. First, a statistically significant gain in sensitivity on isolated and unseen-region nodes at 2 false positives per patient on target scanners, using a patient-level bootstrap. Second, typical-node sensitivity no more than two points below DGST.

Why not just…?

Why not just fine-tune the in-domain lymph node foundation model on a few target scans? This is the most serious threat. DGST already transfers across anatomical regions well. However, a handful of labeled target scans contains few or no isolated nodes, so fine-tuning mostly recalibrates canonical-node behavior. The k-shot arm of the comparison tests this directly. If DGST closes the atypical-node gap with 10 labels, we will report that per-scanner alignment adds nothing in the labeled regime, and the contribution narrows to the label-free setting.

Why not just use heavy intensity, noise and kernel-blur augmentation? For CT this may be enough. Reconstruction kernels are approximately sharpening or smoothing filters, and nnU-Net-style augmentation already covers some of that. We include a strong-augmentation nnU-Net that simulates kernel and noise variation. If it matches our method on atypical nodes, the honest conclusion is that augmentation suffices, and the stratified benchmark becomes the main contribution.

Why not just harmonize images across scanners first? The LMSB study shows that pixel translation methods distort fine structure unless guided by dense anatomy, and small low-contrast nodes are the structure most at risk. No reported cross-vendor CT harmonization result was found on Lacuna. We run a CycleGAN-style harmonizer and measure node-level sensitivity before and after harmonization, which directly tests whether harmonization erases nodes.

Why not just test-time train the network on the target scans? Ayzenberg et al. use test-time training. Without labels, though, its objective can adapt toward whatever dominates the scan, which is normal anatomy, and it gives no guarantee about rare nodes. We include it as a baseline at matched compute.

Proposed Experiments

The first decisive check is the hypothesis itself, measured as the atypical gap. This is the cross-scanner drop in sensitivity for isolated and unseen-region nodes minus the drop for typical nodes, at 2 false positives per patient and stratified by node short-axis size. If that gap is not positive for the naive and station-aware baselines, the motivation fails, and we say so.

The second check is whether tissue-anchored alignment reduces the atypical gap more than all-patch alignment and more than strong augmentation, at equal compute.

The third check settles the transfer question. We compare post-alignment residual $W_2$ and atypical-node sensitivity for DINOv2 against the RADCURE encoder. Natural-image features earn their place only if they align with lower residual and give higher atypical sensitivity.

The final check is the thyroid cohort. We ask whether image-based skip scores add to the six clinical predictors of Zhu et al. on validation AUC, starting from the reported 0.721. We test this separately on scans from scanners not seen in training.