COVID lung infection on CT

Slice viewer: the full series

All 255 axial slices of the RICORD series, segmented with the pipeline below. Drag the slider or use the arrow keys to move through the scan; toggle the infection overlay.

← →

CT only Infection overlay

Slice 128 of 255

-% of lung infected

95% CI -

Slice index	-
Lung voxels	-
Infected voxels	-
Slice HU min / mean / max	-

Precomputed with the segmentation pipeline described below; slice masks use the same fixed metric (infected voxels inside the lung mask). The interval is a Wilson score interval for the infected-voxel fraction, computed in your browser from the slice counts.

Pipeline run, step by step

The production pipeline in the repo (pipeline/) walks a DICOM tree, converts each series to NIfTI, and writes a versioned Parquet manifest that every downstream step reads. This is the actual run behind this page, stage by stage, with the real values from its artifacts. Click a stage to inspect it.

How the pipeline works

The segmentation runs in three stages, all with fixed, interpretable parameters:

1. Lung mask. Keep voxels between -950 and -300 HU, fill holes, compute a distance map, run a watershed split, keep the largest region, and smooth with a median filter of radius 5. This is the standard cheap lung segmentation; it can leak into bowel gas or miss collapsed lung.
2. COVID lesion mask. Inside the lung mask, keep voxels between -700 and -200 HU, then median-filter with radius 2 to drop specks. This band covers ground-glass opacity, the hallmark COVID pattern on CT, but it also catches vessels, partial-volume edges, and motion blur. A threshold cannot tell those apart.
3. Quantification. Infection percentage = infected voxels ÷ lung voxels × 100, per slice and for the whole scan. The original version of this project divided by the wrong denominator and reported values over 100%; that bug is fixed here and documented in the repo.

On the full scan (255 slices, one patient) the fixed method finds 9.69% of lung volume affected. Full analysis, figures, and code: github.com/nepalanurag/Biomedical-Imaging-Analysis.

Data: MIDRC-RICORD-1A (RSNA International COVID-19 Open Radiology Database), via The Cancer Imaging Archive, CC BY-NC 4.0. One COVID-positive chest CT series used here.

Try your own image

Upload a lung-window CT slice (grayscale PNG or JPEG). The page shows it and gives a rough on-screen estimate by inverting the standard lung window back to approximate Hounsfield units. This is approximate: there is no lung mask in the browser, so treat it as a sketch, not a measurement. The viewer above carries the real segmented results.

Validated against radiologist annotations

The MIDRC-RICORD-1A collection ships radiologist-drawn annotations of infectious opacities (MD.ai format, annotator Team 1). 142 of the 255 slices in this series carry 346 annotation polygons. Scoring the pipeline's prediction against them:

Dice measures overlap between the method's mask and the radiologist's drawings: 1.0 is perfect agreement, 0 is none. 0.71 is a solid result for plain HU thresholding with no learned model. The annotations also drove a real fix: the lung mask was losing infection near the pleura, and repairing it with morphological closing lifted Dice from 0.67 to 0.71.

One series, one annotator team: these numbers describe this scan, not COVID CT in general. Reproduce them with python evaluate_annotations.py in the repo (needs the annotations JSON from the TCIA MIDRC-RICORD-1A collection page; see the main README).

Across patients

Infection percentages were computed for every scan series in the study set. Series reporting above 100% are an artifact of the old denominator bug and are dropped, matching the original analysis. That leaves 206 series from 109 patients.

A note on the numbers: the chart below shows the legacy study table, where series above 100% were dropped to match the original analysis. The repo now also carries a curated version of that table (infection_quantification_by_subject_curated.csv): the new series-eligibility rule keeps 154 axial diagnostic series and excludes 75 scout, localizer, and reformat series. Four kept rows still carry values above 100% from the pre-consolidation numerator; they are kept as-is and flagged for recomputation on the next refresh, not silently dropped. Details in the curation panel under Production pipeline.

Distribution of infection percentage across the 206 kept series. Data: infection_quantification_by_subject.csv in the repo.

Production pipeline

The method above now runs inside a production ingestion pipeline in the repo (pipeline/). It walks a DICOM tree, converts each series to NIfTI, and writes a versioned Parquet manifest that every downstream step reads. Validation gates check the schema contract, the de-identification pattern, and the series-eligibility rule before anything is quantified. Rejected series are quarantined with a machine-readable reason instead of being silently dropped, and every refresh is stamped with a timestamp, code hash, and git SHA. The repo carries 30 tests, DVC stages, a Dockerfile, and CI for the pipeline.

Validation on the real series

End-to-end run on the 255-slice RICORD series used in this demo:

Whole-scan result: 9.69% of lung volume infected, 95% CI [9.67, 9.71], from 682,603 infected voxels out of 7,043,692 lung voxels.

Series curation

The eligibility rule was applied to the legacy 229-row study table. 75 series were excluded as non-diagnostic, leaving 154 axial diagnostic series:

Curated table and curation log: infection_quantification_by_subject_curated.csv and analysis/curation_log.csv in the repo.

HU-band sensitivity sweep

The pipeline re-ran quantification over 13 threshold configurations (18.6 minutes wall time): the project-standard bands plus one-at-a-time variations of each lung and infection band edge. The headline 9.69% sits inside a sweep range of 0.24% to 14.47%. Click a bar to inspect that configuration.

The infection band's upper edge is the dominant lever: raising it pulls dense consolidation into the mask, lowering it restricts the mask to ground glass. The lung band edges move the denominator and the mask together, so their effect is smaller, except at the extreme. Setting the lung lower bound to −1000 collapses the metric to 0.24%, because background padding floods the lung mask. That configuration is a documented failure mode, not a plausible alternative.

Every configuration is a full pipeline run with its code hash stamped; cost scales with configurations × series × segmentation time. Sweep code: pipeline/ai_sweep.py in the repo.

AI second reader

A vision model (Gemini 2.5 Flash) was shown the same four infection-overlay slices as the earlier dry-run study and asked to judge four known failure modes of threshold segmentation. Its reads are uncalibrated hypotheses, not ground truth, and they were checked against the analyst's own written takeaways:

Agreement with the analyst's takeaways: the method-level points (T1–T5, T7) cannot be judged from overlays either way; T6 (documented disagreement modes versus radiologist polygons) was partially reproduced: the model flagged the missed consolidation and vessel misclassification but did not reproduce the partial-volume edge finding. Full report: pipeline/data/disagreement_report_vlm.md in the repo.

Analyze your own scan

Upload a zip of DICOM files from a chest CT series and the real pipeline (ingest, validation, eligibility, quantification) runs on it in the cloud. A full 255-slice series takes a few minutes; you get the infection percentage with its confidence interval, the validation outcome, and infection overlay images back.