Code and documentation
Interactive demo
Release v0.3.2
Benchmark protocols and measurements

RevCT

A typed decision model for computed tomography (CT).

Rev CT classifies the 18 CT-RATE chest findings in a single image-encoder pass, returning calibrated scores, present/absent decisions and schema-defined records. It achieves state-of-the-art macro-F1 within NVIDIA's published CT-RATE 18-class comparison: 0.6154 macro-F1 and 0.8715 macro-AUROC on the corrected 3,002-volume validation cohort.

The complete model is compact: approximately 147M parameters in a 589 MB checkpoint, including the frozen COLIPRI image encoder. It has no language decoder. With the tested fast processor, warm volume inference averages 2.856 seconds on an A100 80 GB, 7.63× faster than the measured NV-Reason-CT 18-question protocol.

Weights and loading metadata are hosted on Hugging Face. The inference package and documentation are on GitHub.

Installation

Volume inference requires Python 3.11+ and PyTorch. Install the inference package:

gh repo clone hcltech-robotics/rev-ct
cd rev-ct
git checkout v0.3.2
python -m pip install '.[inference]'
hf auth login

The supported inference stack uses Transformers 5.6.2, COLIPRI 0.1.3 and dynamic-network-architectures 0.4.3. The image encoder weights are bundled in Rev's checkpoint. COLIPRI supplies the installed architecture and reference processor; a second encoder-weight download is unnecessary.

Quick start

import rev_ct  # registers the installed decision-model architecture
import torch
from transformers import AutoModel, AutoProcessor

repo = "hcltech-robotics/rev-ct"
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
model = AutoModel.from_pretrained(repo).eval().to("cuda")
processor = AutoProcessor.from_pretrained(repo, preprocessing="fast")
inputs = processor(images3d=["chest.nii.gz"], return_tensors="pt")

with torch.inference_mode():
    output = model(**inputs, volume_names=["chest.nii.gz"])

print(output.records[0])
print(output.probabilities)  # calibrated scores, (1, 18)
print(output.decisions)      # trained present/absent rule, (1, 18)

Importing rev_ct registers AutoConfig, AutoModel and AutoProcessor for model_type="decision-model". No trust_remote_code flag is required. Keep the model in fp32 and preserve the head tensor dtypes. Global .half() or .bfloat16() conversions alter the calibrated decision path and are unsupported. Pin the same Hub revision for both objects in deployed applications.

Reference preprocessing remains the processor default. The explicit fast option uses the tested SimpleITK path and falls back to the reference processor for unsupported geometry. save_pretrained() stores weights and metadata, reloadable with the installed package, including offline with local_files_only=True.

Licence/terms of use

Rev code and learned heads use Apache-2.0. The bundled Microsoft COLIPRI image encoder retains its MIT licence and notice in COLIPRI_LICENSE. CT-RATE is gated and has separate CC BY-NC-SA 4.0 dataset terms. This release does not redistribute CT volumes. See LICENSE, NOTICE and upstream terms.

Use case

Chest-CT classification research, reproducible evaluation and applications requiring a fixed vocabulary of typed finding decisions. The model has not been clinically validated. It cannot establish a complete normal examination or generate a radiology report.

Model architecture

Architecture type: Decision model, frozen 3D transformer encoder with calibrated linear heads.
Network architecture: COLIPRI-CRM pooled, projected 768-dimensional image features; training standardisation; 18 class-balanced logistic regressions; training-only F1 decision biases; per-label isotonic score calibration.
Task: Multi-label chest-CT classification with typed outputs.
Base model: microsoft/colipri, image encoder only.
Model type: decision-model.
Model size: Approximately 147M parameters for the complete model.
Checkpoint: One complete model.safetensors, 588,703,212 bytes.

The unused COLIPRI text encoder is omitted. The bundled encoder tensors match the pinned upstream image checkpoint exactly. Decisions threshold the bias-adjusted logits at zero, equivalent to a fixed 0.5 sigmoid threshold. Calibrated scores are reported separately and need not cross 0.5 at the decision boundary.

Input

Input type: Image.
Input format: 3D NIfTI (.nii or .nii.gz) with intensities in Hounsfield units and valid physical spacing/orientation.
Preprocessing: RAS reorientation, 2 mm isotropic resampling, HU clamp to [-1000, 1000], scaling to [-1, 1], centred 192³ crop/pad.
Alternative input: Pooled, projected, unnormalised 768-dimensional features from the pinned COLIPRI encoder, rounded through fp16 as in training storage.

The reference processor uses TorchIO and centre cropping. It does not perform an anatomical lung crop. Convert DICOM series to a correctly oriented HU NIfTI before inference. For CT-RATE reproduction, use dataset/valid_fixed/ rather than the uncorrected validation volumes.

Output

Output type: Typed multi-label decisions.
Output format: PyTorch tensors and strict JSON-compatible rev-1.0 records.
Output fields: Per-finding decision, calibrated score, decision confidence; three highest-scoring positive findings; mean-score burden and positive finding count.

RevOutput includes logits, raw_probabilities, probabilities, decisions and records. Confidence is the calibrated score for a present decision and one minus that score for an absent decision. The score can disagree with the separately trained decision threshold. The burden summary is a mean score, not a validated clinical severity measure. The package also provides a Clef-shaped typed request interface for questions about the trained finding vocabulary.

Finding vocabulary

Medical material; Arterial wall calcification; Cardiomegaly; Pericardial effusion; Coronary artery wall calcification; Hiatal hernia; Lymphadenopathy; Emphysema; Atelectasis; Lung nodule; Lung opacity; Pulmonary fibrotic sequela; Pleural effusion; Mosaic attenuation pattern; Peribronchial thickening; Consolidation; Bronchiectasis; Interlobular septal thickening.

Software integration

Runtime engines: PyTorch and Transformers, with NumPy/SciPy for the reference calibrated heads.
Operating system tested: Linux.
GPU tested: NVIDIA A100-SXM4-80GB (Ampere).
CPU support: Head inference and full checkpoint loading are verified. CPU volume inference is supported but its speed is not reported here.

Code, CLI, dependency specification, tests, complete inference documentation and benchmark scripts are in the GitHub release. Standard inference reads safetensors and JSON only. It does not load pickle, joblib or scikit-learn objects.

Model version

0.3.2. Complete safetensors checkpoint with installed Transformers decision-model inference.

Training and evaluation datasets

Training dataset

Data modality: Chest CT and report-derived labels.
Dataset: CT-RATE training split, with the reported final head-training frame of 46,144 chest volumes.
Labelling: Official RadBERT labels overwritten where clean gpt-oss-120b report labels are available. Parsing failures retain the official labels.
Encoder pretraining: CT-RATE training scans and NLST; the encoder is frozen for Rev head training.

The heads use LogisticRegression(C=1.0, class_weight="balanced", max_iter=2000). Five shuffled training folds with random_state=0 supply out-of-fold probabilities for decision bias selection and isotonic calibration. The final heads are refitted on the training frame. Biases and calibration use training OOF probabilities only. Validation targets retain the official RadBERT labels.

Evaluation dataset

Dataset: Official CT-RATE validation, 3,039 volumes.
Corrected cohort: 3,002 volumes after the 37 official no-chest exclusions, matching NV-Reason-CT's reported cohort.
Targets: Official RadBERT validation labels for all 18 findings.
Metric protocol: Unweighted macro positive-class F1 at the fixed trained decision threshold; macro AUROC on raw sigmoid probabilities.

Benchmarks

CT-RATE classification, 18 labels. n/r = not reported. Rev uses supervised CT-RATE heads and training-OOF decision biases. NVIDIA reports direct Yes/No prompts without a classification head or task-specific adaptation.

Rev validation results

Accuracy is evaluated on cached validation embeddings. All 54,702 decisions match the source model, with maximum calibrated-score error 5.96e-8. Volume inference is checked on 64 scans: all 1,152 decisions match cached decisions. GPU save/reload preserves all 144 scores and decisions in an eight-scan check.

Inference speed

Test hardware: Hugging Face Jobs, A100-SXM4 80 GB, 12 allocated vCPU.
Sample: 64 distinct CT-RATE patients, 16 each of four acquisition strata.
Timing scope: Warm local NIfTI to CPU-visible output, including preprocessing and postprocessing, excluding downloads and model loading.
Protocol: Batch one, one timed pass per scan/protocol, synchronised CUDA, one Torch/SimpleITK CPU thread, TF32 disabled.

The same 64 scans are used throughout. NV uses verified fast kernels, thinking disabled and greedy decoding. The measured prompts are speed protocols, not a reproduction of NVIDIA's reported classification accuracy. The list protocol returned zero valid requested 18-key Boolean JSON mappings out of 64 responses. Pillar has no fitted CT-RATE head in this timing study. Its feature timings and NV's free-text timings cover different output tasks.

Fast preprocessing reduces mean CPU preprocessing from 6.884 to 1.961 s (3.51×) and total serial latency by 2.72×, with pixel-identical inputs on all 64 scans. Four preprocessing workers reach 0.701 scans/s per GPU, including pool startup for each 16-scan stream, with identical scores and decisions. These figures are not the combined throughput of four GPUs. The acquisition-balanced sample oversamples uncommon protocols and is not an estimate of a hospital's case distribution.

Limitations

The labels are report-extracted weak supervision. The 18 findings do not cover all pathology. Training out-of-fold splits are at volume level; related reconstructions may cross folds. A baseline subgroup audit identified reduced performance on small-FOV acquisitions; the final model has not been separately evaluated in those subgroups.

Site-specific prevalence, coverage and acquisition changes require separate evaluation of decisions and score calibration. The model is for research and has not been clinically validated.

References and provenance

• COLIPRI, image checkpoint 892e42ec363f6c561b89c459d60fcd95ac1a99bf.
• CT-RATE, revision deeca4d89e9f978d4d1bccd88a55071ddbb146bb.
• NV-Reason-CT, comparison revision 386b93e034983f6c1fc841a43833a1b6a0cd9c13.
• Pillar-0 ChestCT, revision b9dfd833a947744382d4f1bd73b5dd72fcb58b34.
• Pillar preprocessing, with the chest fine-tuning recipe.
• Accuracy reproduction and timing data.

Citation

Rev was developed at the HCLTech Robotics Intelligence CoE. If you use RevCT in your work, cite the software release and report the version and checkpoint revision used.

von Csefalvay, C., & Foldi, T. (2026). RevCT: A typed decision model for computed tomography (CT). (Version 0.3.2) [Software]. HCLTech Robotics Intelligence CoE. https://doi.org/10.57967/hf/10861

@software{voncsefalvay2026revct,
  author  = {von Csefalvay, Chris and Foldi, Tamas},
  title   = {RevCT: A typed decision model for computed tomography (CT)},
  year    = {2026},
  version = {0.3.2},
  doi     = {10.57967/hf/10861},
  url     = {https://github.com/hcltech-robotics/rev-ct}
}