AI Analysis of a Low-Dose Lung CT (LDCT) Screening Study

A transparent demonstration of what a general-purpose multimodal LLM (ChatGPT / Claude) can and cannot do with a real DICOM imaging dataset.

Author: Analysis performed by Claude Opus 4.7 (Anthropic), with cross-reference to an independent ChatGPT (GPT-4 family) analysis of the same study Report date: 2026-04-21 Study date: 2026-04-02 Patient identifiers: De-identified for this report (raw summary.txt on the source disc contains PII; it was deliberately removed here)

1. Purpose & Audience

This report is intended for clinicians and radiology professionals who want an honest look at how well a general-purpose multimodal LLM performs on routine screening CT data, and where its limits are. The study used is a real low-dose lung CT (LDCT) acquired from a free community screening campaign in Hong Kong.

This is not a diagnostic report. It is a technical demonstration.

2. Study Metadata

Extracted directly from DICOM headers via pydicom.

Field Value
Modality CT
Body part examined CHEST
Study description Thorax^CUHK_LUNG_LDCT_V2 (Adult)
Protocol CUHK_LUNG_LDCT_V2 (Chinese University of Hong Kong low-dose lung CT v2)
Scanner manufacturer SIEMENS
Scanner model SOMATOM Drive
Institution Hong Kong Health Check
Study date/time 2026-04-02, 16:08:08 local
Series count 4
Total instances ~1,967 across all series
Total data volume ~1.1 GB on the source DVD

2.1 Series breakdown

# Folder Description Slices Thickness In-plane pixel Kernel Image type
1 17310000 Topogram 1.0 Tr20 500 — 1.00 × 1.00 mm — ORIGINAL, PRIMARY, LOCALIZER, CT_SOM5 TOP
2 17310001 LungCARE_V2_lung 0.6 Bl57 1 500 0.60 mm 0.71 × 0.71 mm Bl57 (lung) ORIGINAL, PRIMARY, AXIAL, CT_SOM5 SPI
3 17310002 LungCARE_V2_BONE 0.6 Br57 1 500 0.60 mm 0.71 × 0.71 mm Br57 (bone) ORIGINAL, PRIMARY, AXIAL, CT_SOM5 SPI
4 17310003 LungCARE_V2_mediast 3.0 MPR cor 467 3.00 mm 0.99 × 0.99 mm — DERIVED, PRIMARY, LOCALIZER, CT_SOM5 MPR

3. Methodology

3.1 Pipeline architecture

The AI does not read DICOM natively. A preprocessing pipeline renders images that the LLM’s general-purpose vision layer can consume:

  ┌─────────────┐     ┌──────────┐     ┌─────────────────────────┐
  │ DICOM files │────▶│ pydicom  │────▶│ structured metadata     │──┐
  │ (~1,967     │     │ (I/O lib)│     │ (tags as text)          │  │
  │  slices)    │     └──────────┘     └─────────────────────────┘  │
  │             │          │                                        │
  │             │          └───▶ raw pixel arrays                   │
  │             │                       │                           │
  │             │                       ▼                           │
  │             │          ┌─────────────────────────┐              │
  │             │          │ numpy + PIL             │              │
  │             │          │ • HU rescale (slope/    │              │
  │             │          │   intercept)            │              │
  │             │          │ • windowing (WL/WW)     │              │
  │             │          │ • normalise to uint8    │              │
  │             │          └─────────────────────────┘              │
  │             │                       │                           │
  │             │                       ▼                           │
  │             │                 ┌──────────┐                      │
  │             │                 │   PNG    │                      │
  │             │                 └──────────┘                      │
  │             │                       │                           │
  └─────────────┘                       ▼                           │
                           ┌────────────────────────────┐           │
                           │ LLM multimodal vision      │◀──────────┘
                           │ (general-purpose, not      │
                           │  radiology-specialised)    │
                           └────────────────────────────┘
                                        │
                                        ▼
                              natural-language findings

3.2 Libraries used

Component Library Role
DICOM parsing pydicom 3.0.2 Reads DICOM headers and pixel data. No interpretation.
Array processing numpy HU rescale, windowing math, normalisation
Image encoding Pillow (PIL) Save uint8 arrays as PNG
Visual analysis LLM vision (Claude 4.7 / GPT-4) Pattern interpretation of rendered PNGs

3.3 Windowing preset

Lung-window reconstructions used:

  • Window Level (WL): −600 HU
  • Window Width (WW): 1500 HU
  • Applied after RescaleSlope × pixel + RescaleIntercept conversion to HU.

Same reasoning applied implicitly by the viewer for Series 3 (bone window) via its saved reconstruction kernel Br57.

3.4 Sampling strategy

Sample Coverage
1 middle slice per series Series 1, 3, 4 (one each) — coarse orientation
9 evenly-spaced slices Series 2 only, at indices 49/99/149/199/249/299/349/399/449 of 500 (~10/20/…/90%), sorted by ImagePositionPatient z-axis so the stack runs base → apex
Total sampled 12 PNG slices out of 1,967 total instances (~0.6%)

Out of the full volume that a radiologist would scroll through in a few seconds, this report examines 0.6% of the data. That single fact sets a hard ceiling on the sensitivity of this analysis.

3.5 Reproducibility

Scripts and outputs are preserved:

Artefact Path
Metadata extractor ~/Downloads/DICOM_for_ChatGPT/dicom_study_summary.py
Slice extractor ~/Downloads/DICOM_for_ChatGPT/extract_slices.py
Python environment ~/Downloads/DICOM_for_ChatGPT/.venv/
Extracted 4-series middle slices ~/Downloads/dicom_output/previews/*.png
Series 2 base→apex walkthrough ~/Downloads/dicom_output/series2_walkthrough/s2_01..09.png
Machine-readable summary ~/Downloads/dicom_output/summary.json

4. AI Visual Findings

All observations below are the LLM’s own reading of the rendered PNGs, with no access to priors, no scrolling, no measurement tools, no windowing changes.

4.1 Orientation series

Series Level sampled Observation
Series 1 (topogram) Middle Axial soft-tissue-window slice showing heart silhouette and flanking lung fields. Served scout/planning role; not used for diagnostic reading.
Series 3 (bone window, middle slice) Mid-thorax Ribs continuous and intact at this level. Vertebral body and spinous process normal. Sternum visible anteriorly. No obvious lytic or sclerotic lesions at the sampled level.
Series 4 (coronal MPR, middle slice) Mid-coronal plane Spine stacks cleanly midline. Symmetric chest outline. Dark vertical bars at image edges are scanner table / gantry artefacts, not anatomy.

4.2 Series 2 — lung-window base → apex walkthrough (9 slices)

Slice Position AI observation
1 10% (idx 49) — upper abdomen Mostly sub-diaphragmatic; liver dominates R, stomach/spleen on L. Lung-base slivers visible at top. Left upper-abdomen dark pockets = gastric/bowel gas (normal). No lung-base findings to flag.
2 20% (idx 99) — lung bases + liver dome Lower-lobe crescents visible around diaphragm. Right hemidiaphragm higher than left (normal liver effect). Normal vascular markings at bases. No focal consolidation, no pleural effusion visible, no nodules.
3 30% (idx 149) — lower thorax / heart Heart central; lungs symmetric and well-aerated flanking the heart. Vessels taper smoothly from hila outward. No focal densities.
4 40% (idx 199) — mid-chest Heart chambers differentiated internally. Lung fields clean. Normal vascular branching. No asymmetry, no lobar collapse pattern.
5 50% (idx 249) — carina region Main airways (carina, main bronchi) visible centrally. Hilar pulmonary arteries + bronchi symmetric left/right. No hilar enlargement or mass effect. Parenchyma clean.
6 60% (idx 299) — upper-mid lung Trachea central. Upper-lobe lung with normal vascular pattern. No nodules, no focal opacities.
7 70% (idx 349) — upper lung Trachea clearly central. Upper lobes symmetric, well-aerated. Faint streak patterns near dense structures interpreted as routine beam-hardening/motion artefacts.
8 80% (idx 399) — high upper lung Lung cross-section tapering toward apex. Bony thorax (ribs) prominent. Parenchyma dark and clean; no apical scarring or pleural thickening flagged.
9 90% (idx 449) — apex / thoracic inlet Small apical lung regions. Dominated by thoracic-inlet structures: trachea, neck/shoulder soft tissue, clavicles, shoulder joints. Both lung apices appear clear; no apical masses, cavities, or pleural thickening flagged.

4.3 Aggregate read (Series 2, 9-slice sampling)

  • Aeration: uniformly dark (well-aerated) from base to apex, bilaterally; symmetric at every sampled level
  • Vascular pattern: normal-looking branching, smooth tapering from hila to periphery at every level
  • Airways: trachea and main bronchi patent and central throughout
  • No AI-flagged findings: no clear solid nodules > ~5–8 mm, no obvious consolidation, no lobar collapse, no bullae/cystic change, no honeycomb fibrosis pattern, no obvious pleural effusion, no apical scarring

This is consistent with the negative binary screening result returned by the campaign. It is not independent confirmation of it.

5. Known Limitations of This Approach

# Limitation Clinical implication
1 8-bit PNG input LLM sees ~256 gray levels vs DICOM’s ~4096. Subtle density differences (e.g., ground-glass opacity) are compressed.
2 No volumetric viewing LLM sees isolated static slices. Radiologists scroll 500+ slices continuously to track structures across planes.
3 Sampled (~0.6%) coverage A 6 mm nodule between sample points is invisible.
4 Single window per pass Radiologists switch windows fluidly. LLM gets one pre-rendered window per request.
5 No priors / serial comparison Growth is the dominant signal in screening. LLM has none.
6 No measurement tools LLM estimates sizes by appearance; cannot quantify in millimetres against a calibrated ruler.
7 No radiology-specific supervision Model was trained on general web imagery + broad text; it has seen medical images but not at the depth/rigour of a radiology-trained CNN.
8 No regulatory clearance Not CE-marked, not FDA-cleared, not a certified medical device.
9 Hallucination risk LLMs can confidently confabulate findings. Cross-checking between models (ChatGPT vs Claude) and against ground truth is essential.

6. Context: General-Purpose LLM vs Specialist Radiology AI

Aspect ChatGPT / Claude (this report) Specialist radiology AI (e.g., Aidoc, Lunit INSIGHT CXR, Siemens AI-Rad Companion)
Input format Rendered 8-bit PNG (via pydicom pipeline) Native DICOM, full bit depth
Model architecture General multimodal transformer Task-specific CNN / transformer trained on labelled radiology studies
Training data Web-scale text + images (incl. medical) Millions of radiologist-labelled studies
Output Natural-language description Structured findings + measured confidence + bounding boxes
Regulatory status Not a medical device CE / FDA cleared for specific indications
Typical use Orientation, education, patient literacy Triage, worklist prioritisation, second-reader
Sensitivity to small nodules Low-to-moderate at best High (target-designed)
Hallucination risk Present Very low (task-bounded)

The two are complementary, not interchangeable. General LLMs make medical imaging accessible to laypeople and educators; specialist AI provides clinically useful automated reads under regulatory supervision.

7. What a General-Purpose LLM Is Useful For Here

Despite the limits above, this workflow has genuine value:

  1. Patient-facing education. A participant in a free screening programme who gets only a binary result can get a meaningful plain-language orientation to what was scanned and what was looked for.
  2. DICOM/data literacy. The metadata-level explanation (series, kernels, windows, HU) teaches the imaging pipeline itself, useful for medical students, data scientists, and technically-minded patients.
  3. Triage of the user’s own curiosity. “Should I ask for a follow-up?” kinds of questions can be framed against what the AI thinks is gross-level normal vs abnormal.
  4. Research / demonstration. Transparent comparisons like this one inform what generalist AI can and cannot contribute to radiology workflows.

8. Disclaimer

This report was produced by a general-purpose LLM without radiological training, without regulatory clearance, and without access to priors. It does not constitute a clinical diagnosis, a second-reader opinion, or any form of medical advice. Any concern about a patient’s imaging must be directed to a qualified radiologist reading the original DICOM study in a certified workstation.

Appendix A — Reproducibility commands

# Metadata + 3 representative PNGs per series
python dicom_study_summary.py \
  --root ~/Downloads/DVD_2604020417 \
  --out  ~/Downloads/dicom_output

# 9 evenly-spaced lung-window slices from Series 2
python extract_slices.py \
  --series ~/Downloads/DVD_2604020417/DICOM/26040209/17310001 \
  --out    ~/Downloads/dicom_output/series2_walkthrough \
  --n 9 --window lung --prefix s2

Appendix B — Windowing reference used by the extractor

Preset WL (HU) WW (HU) Used for
lung −600 1500 Parenchymal evaluation
mediastinum 40 400 Vessels, soft tissue, nodes
bone 500 2000 Osseous detail
soft 40 400 Generic soft tissue
brain 40 80 CNS (not applicable here)