PanTrack: Pancreatic CT Benchmark

• PanTrack is a longitudinal pancreatic cancer CT benchmark featuring consistent, expert lesion-level annotations across multiple timepoints.
• It provides a rigorous out-of-distribution test for retrieval and delineation pipelines in both fully automatic and clinician-verified tracking settings.
• The benchmark supports robust cross-domain validation using metrics like DSC, NSD, and LDR over 45 patients and 161 CT scans.

PanTrack is a public longitudinal pancreatic cancer CT benchmark introduced in “Exploiting Longitudinal Context in Clinician-Verified Interactive Lesion Tracking” (Kirchhoff et al., 22 May 2026). It is designed for longitudinal lesion tracking in serial CT, with particular emphasis on out-of-distribution generalization: in the originating study, PanTrack is used not for training or model selection, but as an entirely unseen evaluation set for methods developed on a different longitudinal lesion domain. Within that framework, PanTrack serves two coupled purposes. First, it addresses the scarcity of public datasets with consistent, lesion-level instance annotations across multiple timepoints. Second, it provides a stress test for workflows that separate lesion retrieval from lesion delineation while still exploiting baseline lesion appearance as a longitudinal prior (Kirchhoff et al., 22 May 2026).

1. Origin and intended role

PanTrack was introduced as a longitudinal pancreatic cancer benchmark for oncologic follow-up in serial CT. The underlying problem is to identify a lesion in a baseline scan, retrieve the corresponding lesion in a later scan, and delineate its volume for response assessment. The benchmark is therefore embedded in a broader distinction between retrieval of the corresponding lesion and delineation of that lesion once correspondence has been established (Kirchhoff et al., 22 May 2026).

Its role in the originating work is explicitly out-of-distribution. Model development is performed exclusively on the longitudinal autoPET IV melanoma CT cohort, together with synthetic longitudinal pretraining, and PanTrack is then used as a zero-shot testbed for transfer to a different anatomy and disease presentation. This design makes PanTrack a measure of cross-domain generalization rather than in-domain optimization. A plausible implication is that PanTrack is most informative when a method claims to use longitudinal context in a way that should survive a substantial domain shift, rather than merely memorizing lesion appearance statistics from the training cohort (Kirchhoff et al., 22 May 2026).

The benchmark is also tied to a specific clinical-operational framing. It evaluates both a fully automatic setting, in which follow-up prompting is produced by registration, and a verified setting, in which a clinician accepts or corrects the proposed prompt. PanTrack is thus not a generic segmentation dataset; it is a benchmark for longitudinal retrieval-and-delineation pipelines under both automated and clinician-in-the-loop usage modes (Kirchhoff et al., 22 May 2026).

2. Cohort composition and annotation

PanTrack comprises 45 patients with pancreatic adenocarcinoma and 161 CT examinations. The longitudinal depth ranges from 2 to 11 scans per patient, with a mean of 3.6 scans per patient. All scans are CT acquired in portal-venous phase, at a single institution, and under identical Siemens protocols (Kirchhoff et al., 22 May 2026).

The annotated lesions include all pancreatic lesions across timepoints, and hepatic metastases if present. Annotation was performed by an experienced radiologist with pancreatic imaging expertise, who manually segmented lesions across the longitudinal series. The cohort was constructed to include heterogeneous disease courses: some patients had long-term stable chemotherapy courses, while others showed rapid progression. The dataset therefore supports repeated longitudinal tracking beyond a simple baseline-follow-up pair design (Kirchhoff et al., 22 May 2026).

Attribute	Value	Note
Patients	45	Pancreatic adenocarcinoma
CT examinations	161	Portal-venous phase
Scans per patient	2–11	Mean 3.6
Institution	Single institution	Identical Siemens protocols
Annotation scope	Pancreatic lesions; hepatic metastases if present	Manual longitudinal segmentation
Annotator	Experienced radiologist	Pancreatic imaging expertise

PanTrack is publicly released and is explicitly presented as complementary to autoPET IV. Whereas autoPET IV contains melanoma lesions in whole-body CT, PanTrack focuses on pancreatic adenocarcinoma and associated hepatic metastases. The originating work emphasizes that pancreatic lesions often have fuzzy boundaries and subtle soft-tissue contrast, making the benchmark radiologically distinct and challenging. This suggests that PanTrack is intended not merely as another longitudinal dataset, but as a domain-shift benchmark with clinically meaningful appearance differences (Kirchhoff et al., 22 May 2026).

3. Evaluation design

PanTrack is evaluated under two paradigms: Automatic Tracking and Verified Tracking. In the automatic setting, the system is given only the baseline lesion prompt $p_0$. A registration model, specifically uniGradICON, estimates a deformation field

$\phi: I_0 \to I_t,$

which propagates the baseline point to a follow-up candidate

$\hat{p}_t = \phi(p_0).$

That propagated point becomes the follow-up prompt (Kirchhoff et al., 22 May 2026).

In the verified setting, the clinician inspects the registration-proposed prompt and either accepts or corrects it. For benchmark evaluation, this workflow is simulated by using the ground-truth follow-up lesion centroid as $p_t$. Under this construction, residual errors in verified tracking are interpreted as delineation errors only, because localization errors have been removed by design (Kirchhoff et al., 22 May 2026).

The benchmark reports three metrics in both settings: DSC (Dice Similarity Coefficient), NSD (Normalized Surface Distance), and LDR (Lesion Detection Rate). Results are given as bootstrapped mean $\pm$ standard deviation (Kirchhoff et al., 22 May 2026).

Evaluation mode	Prompt source	Interpretation
Automatic Tracking	Registration-proposed $\hat p_t = \phi(p_0)$	Retrieval and delineation both matter
Verified Tracking	Ground-truth follow-up centroid $p_t$	Delineation error isolated

A central design choice is that PanTrack is completely excluded from any form of training or model selection in the originating paper. No internal train/validation/test split is reported because the benchmark functions there as a held-out external test set. This sharply distinguishes PanTrack from conventional dataset usage and is foundational to its status as an out-of-distribution benchmark (Kirchhoff et al., 22 May 2026).

4. Methodological context: Verified Tracking and longitudinal modeling

PanTrack is most intelligible in relation to the Verified Tracking paradigm proposed in the same work. That paradigm is intended as a middle ground between fully automatic end-to-end trackers, which offer no opportunity to correct silent failures, and decoupled registration-plus-segmentation pipelines, which allow user verification but discard baseline lesion appearance during segmentation (Kirchhoff et al., 22 May 2026).

The method extracts Volumes of Interest (VOIs) around the baseline and follow-up prompts and performs early spatial prompt fusion: $$X_0 = [I_0,\; G(p_0)], \quad X_t = [I_t,\; G(p_t)],$$ where $G(p)$ is a Gaussian heatmap centered at $p$, with $\phi: I_0 \to I_t,$0, rescaled to unit value at the center. These inputs are passed through a shared-weight encoder to obtain multiscale features $\phi: I_0 \to I_t,$1 and $\phi: I_0 \to I_t,$2 (Kirchhoff et al., 22 May 2026).

Longitudinal context is injected through latent temporal difference weighting: $\phi: I_0 \to I_t,$3 This mechanism uses normalized baseline–follow-up feature differences as an attention-like modulation of the follow-up representation. The accompanying ablations argue that naive multi-timepoint fusion can collapse into effectively cross-sectional behavior, whereas difference weighting is needed to exploit longitudinal context effectively (Kirchhoff et al., 22 May 2026).

Synthetic longitudinal pretraining is also central. The model is pretrained on 2,606 CT pairs with synthetic follow-ups simulating tumor growth, tumor shrinkage, and acquisition variability. Prompt simulation uses a 50/50 split between points sampled from the ground-truth lesion mask with probability weighted by $\phi: I_0 \to I_t,$4 and points derived from registered follow-up locations, which may lie outside the lesion. The paper states that synthetic pretraining improves performance by up to 4.5 Dice points over training from scratch, and the ablation study links that gain to the model’s ability to exploit longitudinal information under imperfect prompting (Kirchhoff et al., 22 May 2026).

Although these architectural and training details describe the evaluated method rather than the benchmark itself, they are integral to PanTrack’s function in the literature. PanTrack is the site at which the Verified Tracking formulation is tested under anatomy and disease shift, and the benchmark is explicitly used to determine whether baseline lesion appearance, prompt robustness, and clinician verification remain useful outside the training domain (Kirchhoff et al., 22 May 2026).

5. Reported benchmark results

The originating study evaluates multiple baselines on PanTrack. In automatic tracking, the compared methods are Hering et al., nnInteractive, LesionLocator, and the proposed Verified Tracking model. In verified tracking, the compared methods are SegVol, ULS Model, nnInteractive, and the proposed model (Kirchhoff et al., 22 May 2026).

Setting	Method	DSC	NSD	LDR
Automatic	Hering et al.	$\phi: I_0 \to I_t,$5	$\phi: I_0 \to I_t,$6	$\phi: I_0 \to I_t,$7
Automatic	nnInteractive	$\phi: I_0 \to I_t,$8	$\phi: I_0 \to I_t,$9	$\hat{p}_t = \phi(p_0).$0
Automatic	LesionLocator	$\hat{p}_t = \phi(p_0).$1	$\hat{p}_t = \phi(p_0).$2	$\hat{p}_t = \phi(p_0).$3
Automatic	Proposed method	$\hat{p}_t = \phi(p_0).$4	$\hat{p}_t = \phi(p_0).$5	$\hat{p}_t = \phi(p_0).$6
Verified	SegVol	$\hat{p}_t = \phi(p_0).$7	$\hat{p}_t = \phi(p_0).$8	$\hat{p}_t = \phi(p_0).$9
Verified	ULS Model	$p_t$0	$p_t$1	$p_t$2
Verified	nnInteractive	$p_t$3	$p_t$4	$p_t$5
Verified	Proposed method	$p_t$6	$p_t$7	$p_t$8

In the automatic setting, the proposed method is best on all three metrics. Relative to the strongest automatic baseline in DSC, LesionLocator at 51.7, the gain is +6.5 DSC points. Relative to Hering et al., the gain is +8.3 DSC points. In the verified setting, the proposed method again achieves the best DSC and NSD, while ULS Model attains a slightly higher LDR (95.5 versus 94.7) (Kirchhoff et al., 22 May 2026).

The automatic-versus-verified comparison for the proposed method is numerically modest on PanTrack: verification improves performance by +1.8 DSC, +1.1 NSD, and +1.0 LDR. The authors interpret this as evidence that the model is already robust to registration error on this out-of-distribution dataset. They also emphasize a stronger comparative claim: the model’s automatic PanTrack performance, at 58.2 DSC, exceeds the verified DSC of all competing models (Kirchhoff et al., 22 May 2026).

6. Significance, limitations, and disambiguation

PanTrack fills a specific gap in longitudinal medical imaging: the lack of public resources with multi-timepoint CTs and lesion-level segmentations suitable for combined retrieval-and-delineation evaluation. Its contribution is not simply additional sample count. Rather, it creates a publicly released pancreatic domain benchmark with serial examinations, expert longitudinal annotation, and a deliberately held-out evaluation role. This suggests that PanTrack is particularly useful for benchmarking methods that claim robustness to prompt noise, registration error, or domain shift in longitudinal lesion analysis (Kirchhoff et al., 22 May 2026).

At the same time, several limitations are explicit. PanTrack is used only as an external test benchmark in the originating work, so no internal train/validation/test protocol is reported there. The exact number of derived lesion-pair tracking instances is not given. The paper also does not describe a separate double-reader adjudication workflow for PanTrack analogous to the side-by-side verification used in autoPET IV. These are not defects of the benchmark so much as constraints on what is formally documented in its first presentation (Kirchhoff et al., 22 May 2026).

A common misconception is terminological. In the cited literature, PanTrack denotes the pancreatic longitudinal CT benchmark of (Kirchhoff et al., 22 May 2026). It should not be conflated with PlantTrack, a plant-feature keypoint tracking system for fruits and leaves (Marri et al., 2024); PatchTrack, a Transformer-based multiple object tracking method using frame patches (Chen et al., 2022); or PlanarTrack, a benchmark for planar object tracking by four-corner geometry (Liu et al., 2023). The similarity is lexical rather than conceptual.

In the broader structure of the 2026 work, PanTrack functions as the decisive out-of-distribution test for the Verified Tracking paradigm. Its reported results support the claim that longitudinal baseline context, prompt simulation, and clinician-verifiable prompting can transfer beyond the melanoma domain used for training. Whether that conclusion continues to hold across additional anatomies, acquisition conditions, and lesion taxonomies remains an open empirical question, but PanTrack establishes a concrete benchmark on which such generalization claims can be made precise (Kirchhoff et al., 22 May 2026).