DICOM Packaging Extends NLST CT Annotations to 26,194 Patients

The National Lung Screening Trial offers longitudinal CT imaging and associated clinical data at a scale that is useful for secondary research, but raw scans alone are difficult to reuse for image-analysis studies. Segmentation masks, lesion regions, measurements, and model outputs often arrive in research-specific formats that require specialist tooling and compute to inspect. This paper packages such derived data as DICOM analysis results in the National Cancer Institute Imaging Data Commons (IDC), shifting the contribution from a new prediction model to a reusable annotation and access layer.

Core Contribution

The authors release three DICOM-formatted datasets that complement NLST CT images. IDC NLSTSeg converts expert lesion segmentations and derived measurements for 581 patients; IDC NLSTSybil converts tumor or nodule bounding-box annotations and measurements for 601 patients; and TotalSegmentator-CT-Segmentations supplies volumetric anatomy segmentations plus radiomics features for 26,194 patients.

The new element is the common representation. Rather than distributing masks, boxes, and tabular features as disconnected research files, the work encodes segmentations and Structured Reports alongside the referenced scans using DICOM conventions. That makes the annotation objects usable in tools designed for clinical imaging, while BigQuery tables and dashboards expose their metadata for cohort-level exploration.

Technical Approach

The workflow begins with DICOM files in Google Cloud Storage, ingests them into Google Healthcare, extracts metadata through Google BigQuery, and uses the resulting tables to drive Data Studio dashboards. The OHIF viewer is integrated for direct inspection of images and annotations. Figure 1 lays out this path from cloud storage through DICOM ingestion and queryable tables to dashboards and image review.

For TotalSegmentator, the authors produce anatomy masks and calculate radiomics features for each segment. Their usage notes describe cohort-building heuristics that check segmentation consistency, complete organ coverage, laterality, and anatomical-region volume. The paper also supplies exploratory notebooks and a Streamlit application for examining individual structures.

The NLST-Sybil component pairs lesion bounding boxes with image embeddings from nine foundation models. Those embeddings are used to construct a connectome: an interactive graph whose nodes are patients and whose edges link the five nearest or farthest neighbors in embedding space by cosine similarity. Users can color nodes by patient or clinical attributes, then inspect an associated CT image and closest matches. This is an exploratory retrieval interface, not evidence that the embeddings improve diagnosis.

Results and Analysis

The principal quantitative result is coverage: the TotalSegmentator release spans 26,194 NLST patients, far beyond the 581-patient expert-segmentation cohort and 601-patient Sybil cohort. That scale can make anatomy-derived variables available to researchers who could not practically run segmentation across the collection themselves. The trade-off is that the largest release is generated by TotalSegmentator, whereas the two smaller releases preserve previously shared expert or model-derived annotations.

The paper’s technical validation includes a mapping matrix between TotalSegmentator segment assignments and NLSTSeg expert-assigned tumor locations. The analysis finds overlap for corresponding lung regions, while also showing that extraneous lung regions, cardiac substructures, and ribs may overlap a segmented lesion and disagree with the expert location label. This is a useful warning: anatomical overlap is not equivalent to a clinically correct lesion-location assignment.

Figure 3 illustrates the intended review loop. A dashboard filters an organ and a radiomics feature, identifies a patient with a low middle-right-lung volume, and then uses 3D Slicer to verify that the apparent outlier arose from a cropped CT scan. The workflow therefore supports quality investigation and cohort filtering, but it does not establish a universal threshold for accepting or rejecting automated masks.

Practical Scope

The authors provide DICOM parsing and viewing support through OHIF, a modified DICOMTID1500 plugin in 3D Slicer, and a Google Colab notebook for DICOM Segmentation and Structured Report objects. The released GitHub repository includes creation notebooks, SQL queries, technical-validation notebooks, and usage material. These are concrete components for data engineers and imaging researchers building IDC-based studies.

The evidence supports the claim that the annotations have been standardized, linked to the source images, and made accessible through several interfaces. It is less conclusive on downstream scientific or clinical benefit: the paper presents data infrastructure, qualitative inspection, and consistency checks rather than a benchmark showing improved cancer-risk prediction, lesion detection, or patient outcomes. For users needing broad anatomy-derived covariates, the 26,194-patient release is the consequential part; for lesion-specific work, the smaller expert and Sybil cohorts require attention to their annotation provenance and the observed location disagreements.