PanCT: Pancreatic CT Analysis and Dataset Variants

• PanCT is a collection of heterogeneous pancreatic CT workflows and curated datasets for tumor detection, localization, and segmentation.
• It encompasses methodologies ranging from segmentation-for-classification to multiphase alignment and anatomy-aware conditioning.
• Quantitative benchmarks report high sensitivity and specificity across various PanCT approaches, supporting early diagnosis and treatment planning.

Searching arXiv for papers that define or use “PanCT” in different ways, so the article can reflect the term’s usage precisely. PanCT is a non-unified term in the arXiv literature. In most pancreatic imaging papers, it denotes pancreatic CT analysis or pancreas CT workflows on abdominal contrast-enhanced CT, including tumor detection, localization, and segmentation; in some papers it refers to specific curated pancreatic tumor datasets, such as a dual-phase screening cohort for pancreatic neuroendocrine tumors (PNETs) or a small-lesion segmentation dataset used to evaluate SegDINO; and in still other contexts it is used more loosely, or differently, than a single canonical benchmark name (Zhu et al., 2020, Yang et al., 16 Jun 2026, Li et al., 2 Jul 2025). Taken together, the literature suggests that PanCT is best understood as a family of pancreas-focused CT tasks and datasets rather than a single universally fixed resource.

1. Terminological scope and variant usages

The term has several distinct usages. In pancreatic oncology papers, it commonly designates pancreas CT or pancreatic CT analysis, especially for pancreatic ductal adenocarcinoma (PDAC), PNET, or broader pancreatic tumor segmentation and detection workflows. In the SegDINO paper, PanCT is explicitly “a new CT dataset containing 284 patients with expert-annotated pancreatic tumors,” constructed to benchmark small-lesion segmentation (Yang et al., 16 Jun 2026). In the PanTS paper, the corresponding notion is broader pancreatic CT analysis, with a dataset designed for tumor detection, localization, segmentation, pancreas subregion segmentation, and contextual anatomy modeling (Li et al., 2 Jul 2025).

A concise way to view the literature is as follows.

Usage	Representative source	Brief characterization
Dual-phase PNET screening cohort	(Zhu et al., 2020)	376 abdominal CT cases with arterial and venous phases and voxelwise labels
Small-lesion pancreatic tumor segmentation dataset	(Yang et al., 16 Jun 2026)	284 patients, 243/41 internal train/test split, 2D axial slice modeling
Large-scale pancreatic CT benchmark	(Li et al., 2 Jul 2025)	36,390 CT scans, 145 medical centers, 28 voxel-wise classes
Pan-cancer CT contrast synthesis interpretation	(Chen et al., 22 Jan 2026)	“PanCT” is interpreted as pan-cancer CT contrast synthesis rather than pancreatic CT
Proton computed tomography context	(Pettersen et al., 2020)	“PanCT/pCT” refers to proton computed tomography, not pancreatic CT

This variation matters because a common misconception is to treat PanCT as a single standard public benchmark. The papers do not support that reading. Some works introduce a dataset under that label or a nearby usage, some use the term descriptively for pancreatic CT analysis, and some explicitly reinterpret it in unrelated directions such as pan-cancer CT synthesis or proton CT (Chen et al., 22 Jan 2026, Pettersen et al., 2020).

2. Problem formulations in pancreatic CT

Within pancreatic imaging, PanCT work centers on several related formulations. A foundational line treats detection as segmentation-for-classification: a model first predicts voxelwise tumor maps, then converts those maps into patient-level abnormal versus normal decisions by counting predicted tumor voxels. The PDAC screening paper “Multi-Scale Coarse-to-Fine Segmentation for Screening Pancreatic Ductal Adenocarcinoma” uses this formulation with a tumor-voxel threshold $\tau = 50$ and reports a sensitivity of $94.1\%$ at a specificity of $98.5\%$ (Zhu et al., 2018). The PNET screening paper extends the same logic to dual-phase CT and includes dilated pancreatic duct as an auxiliary cue; a phase is classified as tumor-positive if either $N_{\text{tumor}} \ge 40$ or $N_{\text{duct}} \ge 500$, and the case is abnormal if either arterial or venous phase is abnormal (Zhu et al., 2020).

Another formulation is direct detection or diagnosis from contrast-enhanced CT, sometimes still driven by segmentation and sometimes by lesion-centric classification. The portal-venous PANORAMA challenge solution reformulates early PDAC detection as multi-class segmentation of PDAC-related structures followed by candidate extraction on the PDAC softmax map, reaching an AUROC of $0.9263$ and an AP of $0.7243$ on the testing phase (Liu et al., 13 Mar 2025). The ePAI system targets early and prediagnostic PDAC on routine contrast-enhanced abdominal CT, with a three-stage cascade of anatomical segmentation, lesion localization, and lesion classification (Li et al., 29 Jan 2026).

Multiphase CT is a recurring theme. Radiologists use arterial and venous phases because the phases emphasize different aspects of tumor conspicuity, pancreatic parenchyma, vasculature, and ductal anatomy. The alignment-ensemble paper formalizes PDAC detection on unaligned arterial–venous pairs with early alignment, late alignment, and slow alignment, treating venous as fixed and arterial as moving (Xia et al., 2020). By contrast, some pipelines are explicitly single-phase: the nnU-Net-based PDAC detection framework is trained on portal venous phase CE-CT, and the PANORAMA challenge setting is portal-venous CECT only (Alves et al., 2021, Liu et al., 13 Mar 2025).

A further extension is task expansion beyond tumor presence. The lymph-node metastasis prediction pipeline uses multiphase CT not only to segment and identify lymph nodes but also to derive patient-level nodal metastasis status from lymph-node evidence and tumor-region deep features (Zheng et al., 2023). In radiotherapy, deepPERFECT uses diagnostic CT and planning CT as distinct PanCT roles, synthesizing a planning CT-like image from diagnostic CT to expedite RT planning in locally advanced pancreatic cancer (Hooshangnejad et al., 2023).

3. Datasets and annotation regimes

PanCT-related datasets differ substantially in size, label scope, and intended task. The PNET screening cohort contains $376$ abdominal CT cases with both arterial and venous phases, including $228$ biopsy-proven PNET cases and $148$ normal cases. It provides voxelwise labels with four classes, $94.1\%$0, corresponding to background, normal pancreas, tumor, and dilated pancreatic duct; four expert annotators created the masks in Varian Velocity and a board-certified abdominal radiologist verified them (Zhu et al., 2020).

SegDINO introduces another PanCT dataset for “small lesion segmentation task” evaluation. This dataset contains $94.1\%$1 patients with confirmed pancreatic cancer, with $94.1\%$2 for training and $94.1\%$3 for internal testing. The paper states that 3D CT volumes were converted into 2D axial slices for modeling and that lesions were “independently annotated by two experienced radiologists” (Yang et al., 16 Jun 2026). Unlike the PNET cohort, the paper does not specify scanner vendors, acquisition protocols, contrast phases, slice thickness, in-plane resolution, voxel spacing, or HU windowing ranges (Yang et al., 16 Jun 2026).

The pNET-focused CECT dataset introduced in “An Exceptional Dataset For Rare Pancreatic Tumor Segmentation” contains data from $94.1\%$4 patients and is described as “the first dataset solely dedicated to pNETs.” It includes volumetric 3D CECT studies, multi-phase CECT with at least arterial and venous phases, and 3D tumor segmentation masks cross-validated by expert radiologists (Li et al., 29 Jan 2025).

At a much larger scale, PanTS is a large-scale, multi-institutional CT dataset with $94.1\%$5 CT scans from $94.1\%$6 medical centers across $94.1\%$7 countries. It provides expert-validated, voxel-wise annotations of over $94.1\%$8 anatomical structures across $94.1\%$9 classes, including pancreatic tumors, pancreas head, body, and tail, and $98.5\%$0 surrounding anatomical structures; each scan includes metadata such as patient age, sex, diagnosis, contrast phase, in-plane spacing, slice thickness, and scanner details (Li et al., 2 Jul 2025). The training set contains $98.5\%$1 scans and the test set $98.5\%$2 scans, with the test cohort reserved for third-party evaluation and explicitly described as out-of-distribution relative to training centers (Li et al., 2 Jul 2025).

These annotation regimes encode different assumptions. Some datasets supervise only primary lesion masks, some add pancreas masks, and some incorporate ducts, vessels, pancreas subregions, or broader abdominal anatomy. This suggests that PanCT has evolved from lesion-only labeling toward anatomy-rich annotation, motivated by the observation that pancreatic tumors are small, heterogeneous, and embedded among vessels, ducts, bowel, and other confounders (Alves et al., 2021, Li et al., 2 Jul 2025).

4. Methodological patterns and model architectures

A dominant pattern is encoder–decoder segmentation. Early screening systems used 3D UNet-style backbones with residual connections and deep supervision, often in segmentation-for-classification mode. The PNET screening framework uses a UNet-style encoder–decoder with sum residual connections, deep supervision, voxelwise softmax cross-entropy, $98.5\%$3 patches, sliding-window inference with stride $98.5\%$4, and connected-component post-processing (Zhu et al., 2020). The earlier PDAC screening system trains three volumetric encoder–decoder networks at input scales $98.5\%$5, $98.5\%$6, and $98.5\%$7 and averages their outputs in a coarse-to-fine flowchart (Zhu et al., 2018).

A second pattern is explicit multiphase handling. The alignment-ensemble paper studies three complementary strategies: early alignment by deformable registration in image space, late alignment by deformation in high-level feature space, and slow alignment by iterative multi-level feature registration. Voxelwise predictions from early alignment, late alignment, and slow alignment are combined by majority voting, yielding the best overall performance on both PDAC datasets used in that study (Xia et al., 2020).

A third pattern is anatomy-aware conditioning. The nnUnet_MS model segments pancreas, tumor, and multiple surrounding anatomical structures and outperforms tumor-only or pancreas-plus-tumor variants on external PDAC detection, especially for lesions $98.5\%$8 cm (Alves et al., 2021). PanGuide3D extends this principle into a shared 3D encoder with a pancreas decoder and a tumor decoder conditioned on the probabilistic pancreas map at multiple scales via differentiable soft gating, plus a lightweight Transformer bottleneck (Ma et al., 22 Apr 2026). The lymph-node metastasis work uses organ- and vessel-derived signed distance maps to build a guiding attention map that gates a 3D nnUNet and constrains lymph-node search to anatomically plausible regions (Zheng et al., 2023).

A fourth pattern is self-supervised or foundation-style representation learning combined with lightweight decoders. SegDINO uses a DINOv3-S backbone, Token Pyramid Adaptation (TPA), and Scale-Aware Decoding (SAD), and trains on 2D axial slices resized to $98.5\%$9 with AdamW, learning rate $N_{\text{tumor}} \ge 40$0, weight decay $N_{\text{tumor}} \ge 40$1, cross-entropy loss, batch size $N_{\text{tumor}} \ge 40$2, and $N_{\text{tumor}} \ge 40$3 epochs (Yang et al., 16 Jun 2026).

The field also includes self-learning with partially annotated multi-phase data. A teacher–assistant–student framework trains on $N_{\text{tumor}} \ge 40$4 CT studies from four data sources, uses DEEDS nonrigid registration for inter-phase alignment, refines pseudo labels with an organ/vessel teaching assistant, and trains a final student nnUNet on manual and pseudo annotations jointly (Zhang et al., 2020).

5. Quantitative performance and benchmark behavior

Performance figures vary sharply with task definition. On dual-phase PNET screening, the segmentation-for-classification approach reaches a sensitivity of $N_{\text{tumor}} \ge 40$5 at a specificity of $N_{\text{tumor}} \ge 40$6 after dual-phase OR fusion. On venous single-phase evaluation, the same method reports tumor Dice $N_{\text{tumor}} \ge 40$7, sensitivity $N_{\text{tumor}} \ge 40$8, and specificity $N_{\text{tumor}} \ge 40$9; the direct classification network has the same sensitivity, $N_{\text{duct}} \ge 500$0, but a slightly higher specificity, $N_{\text{duct}} \ge 500$1, at the cost of interpretability (Zhu et al., 2020).

On PDAC screening, the multi-scale coarse-to-fine segmentation-for-classification system reports sensitivity $N_{\text{duct}} \ge 500$2 at specificity $N_{\text{duct}} \ge 500$3 at its main operating point with $N_{\text{duct}} \ge 500$4 tumor voxels (Zhu et al., 2018). On the same 439-case dual-phase PDAC dataset, the later alignment-ensemble work shows that dual-phase alignment improves tumor segmentation beyond single-phase and unaligned baselines; its alignment ensemble reaches tumor DSC $N_{\text{duct}} \ge 500$5, sensitivity $N_{\text{duct}} \ge 500$6, and specificity $N_{\text{duct}} \ge 500$7 on Dataset I, and tumor DSC $N_{\text{duct}} \ge 500$8 plus duct DSC $N_{\text{duct}} \ge 500$9 on Dataset II (Xia et al., 2020).

For pure segmentation benchmarks, SegDINO reports the best PanCT test-set result among its compared baselines, with DSC $0.9263$0 and HD95 $0.9263$1, compared with U-Net at $0.9263$2 and Attention U-Net at $0.9263$3. Its ablation shows that TPA is the dominant contributor on this dataset: the basic model obtains DSC $0.9263$4, TPA only $0.9263$5, SAD only $0.9263$6, and full TPA+SAD $0.9263$7 (Yang et al., 16 Jun 2026).

For cohort-shift analysis, PanGuide3D reports the best tumor performance among its matched baselines both in-cohort and out-of-cohort. On MSD out-of-cohort evaluation it achieves tumor Dice $0.9263$8, tumor sensitivity $0.9263$9, patient sensitivity $0.7243$0, and false-positive tumor volume $0.7243$1 (Ma et al., 22 Apr 2026). PanTS shifts the scale of benchmarking further: nnU-Net trained on MSD-Pancreas yields AUC $0.7243$2 on the PanTS test set, training on PANORAMA yields AUC $0.7243$3, and training on PanTS yields AUC $0.7243$4 on the out-of-distribution PanTS test set (Li et al., 2 Jul 2025).

At the level of early and prediagnostic diagnosis, ePAI reports an internal-test AUC of $0.7243$5, sensitivity $0.7243$6, and specificity $0.7243$7; for small PDAC $0.7243$8 cm, internal sensitivity is $0.7243$9. In the external multicenter diagnostic cohorts totaling $376$0 patients, pooled AUC is $376$1, sensitivity $376$2, and specificity $376$3, while small-PDAC sensitivity is $376$4 (Li et al., 29 Jan 2026). On prediagnostic scans obtained $376$5 to $376$6 months before clinical diagnosis, ePAI detects and localizes future PDAC in $376$7 of $376$8 patients, with median lead time $376$9 days (Li et al., 29 Jan 2026).

6. Clinical significance, ambiguities, and future directions

Across these papers, PanCT is consistently linked to clinical tasks where early localization matters as much as case-level classification. Segmentation-for-classification produces voxel-level tumor or duct maps that radiologists can inspect directly, and several papers explicitly frame this interpretability as a reason to prefer segmentation over black-box classification (Zhu et al., 2020, Zhu et al., 2018). In early PDAC work, indirect signs such as dilated pancreatic duct, duct cutoff, focal atrophy, delayed enhancement, and vessel relationships remain central because lesions may be isoattenuating or only a few millimeters in size (Alves et al., 2021, Li et al., 29 Jan 2026).

Limitations also recur. Many datasets are single-center or internally split; acquisition parameters are often incompletely specified; several methods depend on fixed post-processing thresholds or phase-specific assumptions; and some pipelines, such as SegDINO, explicitly acknowledge the loss of inter-slice context when 3D volumes are converted to 2D axial slices (Yang et al., 16 Jun 2026). Even where large-scale data exist, train/test distribution shift remains substantial: PanTS reports significant differences between train and test in age, spacing, slice thickness, and contrast phases, intentionally creating an out-of-distribution evaluation scenario (Li et al., 2 Jul 2025).

A second ambiguity is conceptual rather than technical. PanCT is sometimes treated as though it were a single dataset or a single task. The literature instead presents a heterogeneous landscape: PNET screening, PDAC screening, tumor segmentation, lymph-node metastasis prediction, pancreas organ segmentation, RT planning synthesis, pan-cancer CT contrast synthesis, and proton CT have all been associated with the label or nearby terminology (Hooshangnejad et al., 2023, Chen et al., 22 Jan 2026, Pettersen et al., 2020). This suggests that any use of the term benefits from explicit disambiguation: whether it refers to pancreatic CT analysis broadly, a particular pancreatic tumor dataset, or a non-pancreatic expansion of the abbreviation.

The most consistent future directions are already visible in the papers themselves: joint segmentation and classification, improved phase-aware fusion, multi-center external validation, larger anatomy-rich annotation, probabilistic or anatomy-conditioned modeling, and integration of radiomics or clinical variables for risk stratification (Zhu et al., 2020, Alves et al., 2021, Li et al., 2 Jul 2025). In that sense, PanCT has come to denote not merely a modality, but a methodological program: exploiting pancreatic CT with increasingly explicit anatomical context, stronger supervision, and more realistic out-of-distribution evaluation to improve detection, localization, and downstream clinical decision support.