---
title: 'LUCID: Unified PET/CT Data for Lung Cancer IGTV'
url: https://www.emergentmind.com/topics/lung-cancer-unified-cross-modal-imaging-dataset-lucid
type: topic
---

# LUCID: Unified PET/CT Data for Lung Cancer IGTV

The Lung-cancer Unified Cross-modal Imaging Dataset (LUCID) is a unified, cross-modal resource curated for lung cancer imaging, designed to integrate anatomical CT and metabolic PET for tasks that require multimodal fusion. In the study that introduces its role in internal gross tumor volume segmentation, the emphasis is on the PET/CT subset, denoted LUCID-PET/CT, which comprises a private internal gross tumor volume cohort collected at Johns Hopkins Hospital. Within that setting, LUCID functions not merely as a repository of images but as the data substrate for transfer learning from gross tumor volume segmentation to the more clinically demanding task of internal gross tumor volume delineation in lung cancer radiation therapy planning [2509.22841].

## 1. Definition and scope of LUCID

LUCID is described as a unified, cross-modal resource curated for lung cancer imaging. Its defining characteristic is the integration of PET and CT, bringing together metabolic and anatomical information within a single data resource. In the reported study, the relevant component is LUCID-PET/CT, a PET/CT subset centered on internal gross tumor volume delineation rather than gross tumor volume delineation alone [2509.22841].

The dataset’s cross-modal formulation is directly relevant to segmentation problems in which neither modality is sufficient in isolation. CT contributes anatomical localization, while PET contributes metabolic uptake patterns. This joint representation is particularly important in lung cancer because tumor boundaries can be difficult to resolve when respiratory motion, complex morphology, and neighboring high-uptake organs degrade boundary clarity in PET, while CT alone may not fully capture the metabolic extent of disease.

A common misconception would be to treat LUCID-PET/CT as a generic PET/CT segmentation benchmark. The study instead frames it as a clinically specific resource for internal gross tumor volume delineation, with explicit downstream relevance to radiation therapy planning. Another important distinction is that the paper discusses only the PET/CT subset in detail; broader claims about other possible LUCID modalities would go beyond the reported evidence.

## 2. Clinical target: IGTV rather than GTV

The study distinguishes gross tumor volume (GTV) from internal gross tumor volume (IGTV). GTV is the visible extent of the tumor on imaging, whereas IGTV extends GTV to account for respiratory motion so that treatment margins reflect the motion envelope in mobile tumors such as lung cancer [2509.22841].

This distinction is central to the dataset’s purpose. PET acquisition occurs over numerous breathing cycles, on the order of 15–30 minutes and 2–5 minutes per bed position. As a result, tumor motion appears as blurring and reduced SUV at lesion margins. The reported challenges include high uptake in adjacent organs such as the heart, motion artifacts, complex morphology, and intrinsically low PET boundary signals. These factors make IGTV boundaries frequently indistinct and therefore difficult to delineate reliably.

The clinical significance is explicit. The planning target volume (PTV) is described as a 5 mm uniform expansion from IGTV to cover delivery uncertainty and setup error. Accurate IGTV delineation is therefore pivotal to reliably covering the true motion envelope. This suggests that LUCID-PET/CT is organized around a radiotherapy-specific contouring problem in which segmentation error can propagate directly into treatment volume definition.

## 3. Cohort composition and acquisition heterogeneity

The LUCID-PET/CT IGTV cohort includes 60 lung cancer patients, yielding 60 IGTV ROIs and 1,067 axial slices in total, with an average of 17.8 ± 8.2 slices per ROI and a range of 9–47 slices. Tumor volumes range from 3.60 to 822.86 cc, with a median of 11.81 cc, and maximum SUV ranges from 3.12 to 23.98, with a mean of 8.25 ± 4.3. Inclusion criteria enforced ROI volume $\geq 3.0$ cc and max SUV $\geq 3.0$ to ensure clinical relevance [2509.22841].

Data acquisition spanned June 2023 to December 2024 on multiple scanners in a multi-vendor setting, approximately 80% GE Medical Systems and 20% Siemens. The dataset aggregates heterogeneous protocols and reconstruction settings, reflecting real-world clinical variability rather than a tightly standardized acquisition regime. CT voxel sizes range from $0.98 \times 0.98 \times 3.00$ mm to $1.52 \times 1.52 \times 5.00$ mm. PET in-plane resolution ranges from $1.65 \times 1.65$ mm to $5.47 \times 5.47$ mm, and PET slice thickness ranges from 2.79 mm to 5.00 mm. Reconstruction methods included OSEM, PSF with TOF, Vue Point High Definition (VPHD), and others.

| Attribute | Reported value | Notes |
|---|---:|---|
| Patients | 60 | Lung cancer cohort |
| IGTV ROIs | 60 | One ROI set per patient |
| Axial slices | 1,067 | Total across cohort |
| Slices per ROI | 17.8 ± 8.2 | Range: 9–47 |
| Tumor volume | 3.60–822.86 cc | Median: 11.81 cc |
| Maximum SUV | 3.12–23.98 | Mean: 8.25 ± 4.3 |

The heterogeneity of scanner vendors, voxel spacing, and reconstruction algorithms is consequential. It means that LUCID-PET/CT is not a homogeneous laboratory dataset but one that explicitly preserves variability encountered in clinical practice. A plausible implication is that methods validated on this cohort are being tested against domain shifts that arise naturally from acquisition diversity.

## 4. Annotation protocol and data standardization

Annotation in LUCID-PET/CT followed a three-stage clinical protocol. First, physicians performed detailed slice-by-slice IGTV delineation on planning CT. Second, PET/CT scans were rigidly registered to the planning CT within spine/lung regions and then deformably registered in the tumor region using Velocity. Third, an independent physician reviewed the results; if inconsistencies were found, a third-stage physician repeated IGTV delineation to ensure consensus quality [2509.22841].

The preprocessing pipeline also reflects the radiotherapy orientation of the dataset. CT preprocessing used the lung CT window $[-1350, 150]$ HU. PET normalization used ROI-specific maximum SUV to mitigate high-uptake confounders. Three-dimensional binary masks were generated from DICOM contour sequences, co-registered with PET/CT slices, and volumes were calculated with anisotropic voxel dimensions. Images were saved as standardized $512 \times 512$ PNGs with comprehensive metadata. For model development, the dataset was split 70%/15%/15% into train/validation/test at the patient level using a fixed random seed.

These steps matter because LUCID-PET/CT is not only heterogeneous at acquisition time; it is also harmonized through a registration and contour-processing workflow that ties PET/CT images to planning CT and consensus annotations. This suggests that the dataset is intended for supervised learning scenarios in which geometric alignment and contour fidelity are critical.

## 5. Role of LUCID in the transfer-learning pipeline

In the reported study, LUCID-PET/CT serves as the target-domain dataset for fine-tuning after pre-training on a large GTV dataset, PCLT20K, which contains 21,930 PET/CT image pairs from 605 patients and is public and primarily dedicated to GTV segmentation. The motivation is explicit: IGTV annotation is limited, so the model is first pre-trained on GTV and then adapted to IGTV using the private LUCID-PET/CT cohort [2509.22841].

The network employed for this transfer-learning setting is a multimodal interactive perception network with Mamba, referred to as CIPA. CIPA uses dual Mamba backbones for PET and CT, with channel-wise rectification (CRM) and dynamic cross-modality interaction (DCIM) to fuse metabolic and anatomical features. Pre-training uses standard 2D inputs, whereas fine-tuning on LUCID-PET/CT extends the architecture to a 2.5D segmentation framework.

The optimization objectives are reported separately for GTV and IGTV:
$$
\mathcal{L}_{GTV} = \mathcal{L}_{Dice}(\mathbf{P}_{GTV}, \mathbf{Y}_{GTV}) + \mathcal{L}_{BCE}(\mathbf{P}_{GTV}, \mathbf{Y}_{GTV}),
$$
$$
\mathcal{L}_{IGTV} = \mathcal{L}_{Dice}(\mathbf{P}_{IGTV}, \mathbf{Y}_{IGTV}) + \mathcal{L}_{BCE}(\mathbf{P}_{IGTV}, \mathbf{Y}_{IGTV}),
$$
with Dice defined as
$$
\mathrm{Dice}(P,G) = \frac{2\lvert P \cap G \rvert}{\lvert P \rvert + \lvert G \rvert}.
$$

Within this framework, LUCID-PET/CT is not just a held-out evaluation set. It is the domain in which the model learns the difference between visible tumor extent and motion-aware internal tumor extent. This suggests that the dataset occupies a transfer-learning role in which clinically specific annotation semantics are acquired only at fine-tuning time.

## 6. Slice-aware 2.5D modeling and the Slice Interaction Module

The study’s fine-tuning regime on LUCID-PET/CT uses a 2.5D segmentation setup rather than single-slice 2D inputs. For each modality, three consecutive slices are concatenated along channels, producing input $\mathbf{X} \in \mathbb{R}^{B \times 3 \times H \times W}$. The rationale is to provide richer local continuity, especially for peripheral slices where PET intensities are weak [2509.22841].

To model inter-slice relationships, the method introduces the Slice Interaction Module (SIM), a lightweight, slice-aware block with three branches. The channel attention branch adaptively weights each slice:
$$
\mathbf{A}_{ch} = \sigma\!\left(\mathbf{W}_2 \cdot \mathrm{ReLU}\!\left(\mathbf{W}_1 \cdot \mathrm{GAP}(\mathbf{X})\right)\right), \quad \mathbf{X}_{ca} = \mathbf{X} \odot \mathbf{A}_{ch},
$$
with reduction ratio $r=2$, parameters $\mathbf{W}_1 \in \mathbb{R}^{C/r \times C}$ and $\mathbf{W}_2 \in \mathbb{R}^{C \times C/r}$.

The spatial attention branch highlights salient spatial regions across slices:
$$
\mathbf{A}_{sp} = \sigma\!\left(\mathrm{BN}\!\left(\mathrm{Conv}_{7 \times 7}(\mathbf{X})\right)\right), \quad \mathbf{X}_{sa} = \mathbf{X} \odot \mathbf{A}_{sp}.
$$

The slice relation branch models inter-slice dependencies via depthwise convolutions:
$$
\mathbf{X}_{rel} = \mathrm{DWConv}_{3 \times 3}\!\left(\mathrm{ReLU}\!\left(\mathrm{BN}\!\left(\mathrm{DWConv}_{1 \times 1}(\mathbf{X})\right)\right)\right).
$$

These branches are fused through a weighted residual connection:
$$
\mathbf{X}_{SIM} = \mathbf{X} + \alpha \cdot \mathbf{X}_{ca} + \beta \cdot \mathbf{X}_{sa} + \gamma \cdot \mathbf{X}_{rel},
$$
with empirically set weights $\alpha=0.3$, $\beta=0.3$, and $\gamma=0.4$.

The stated purpose of SIM is to integrate channel-wise importance, spatial saliency, and slice-to-slice relations so as to mitigate the impact of attenuated PET signals in peripheral slices. In the context of LUCID-PET/CT, this makes the dataset especially suitable for evaluating methods that exploit local axial continuity rather than treating each slice independently.

## 7. Evaluation results, practical implications, and accessibility

Performance on the LUCID-PET/CT test set was assessed with Intersection over Union (IoU), Dice, pixel accuracy (Acc), and 95th percentile Hausdorff Distance (HD95), defined as
$$
\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}, \quad
\mathrm{Dice} = \frac{2|P \cap G|}{|P| + |G|}, \quad
\mathrm{Acc} = \frac{\sum_i \mathbb{I}(p_i = g_i)}{N}, \quad
\mathrm{HD95}(P,G) = \max\{h_{95}(P,G), h_{95}(G,P)\}.
$$
The proposed fine-tuned 2.5D model with SIM achieved $\mathrm{IoU} = 0.438$, $\mathrm{Dice} = 0.609$, $\mathrm{Acc} = 0.846$, and $\mathrm{HD95} = 24.81$. The baseline 2D model without fine-tuning achieved $\mathrm{Dice} = 0.385$ and $\mathrm{IoU} = 0.238$, while the fine-tuned 2.5D baseline without SIM reached $\mathrm{Dice} = 0.532$ and $\mathrm{IoU} = 0.363$. Alternative attention modules, specifically CBAM and Non-Local Block, were reported as less effective than SIM in this slice-aware 2.5D context [2509.22841].

These results connect directly to radiation therapy planning. By leveraging unified PET/CT content and slice-aware modeling, the method is reported to improve IGTV segmentation reliability in clinically challenging scenarios such as weak peripheral PET signal and confounding high uptake in adjacent organs. More precise IGTV contours can support downstream expansion to PTV, for example by uniform 5 mm expansion, on the basis of more accurate motion envelopes. A plausible implication is that this may support robust target coverage while potentially reducing unnecessary irradiation of normal tissue.

The accessibility status of the dataset is also clearly delimited. The LUCID-PET/CT IGTV cohort used in the study is private, de-identified, and collected at Johns Hopkins Hospital. The paper does not provide public access details, licensing terms, or a download link for this subset. By contrast, PCLT20K, the GTV pre-training dataset, is publicly available. This distinction matters for reproducibility: the methodological contribution is documented, but direct external reuse of the LUCID-PET/CT IGTV subset is not enabled by the paper itself.

Source: https://www.emergentmind.com/topics/lung-cancer-unified-cross-modal-imaging-dataset-lucid