AITO-PETCT-MP

About

Editors:

 larsleijten   jacobsc 

Summary

This algorithm predicts the malignancy risk of pulmonary nodules using combined PET/CT imaging. Given a PET/CT scan and the 3D location(s) of one or more nodules, it outputs a calibrated malignancy probability (0–1) per nodule. This model is released for research purposes only and is not intended for clinical use.

Mechanism

Model architecture¶

The algorithm uses an ensemble of five 3D ResNet-18 convolutional neural networks, one per cross-validation fold, trained on PET/CT data from Radboudumc. Predictions are averaged across all five models and subsequently calibrated using Platt scaling (logistic regression) fitted on the development set.

Inputs¶

The algorithm requires three inputs per case:

1. CT image (ct-image) - Format: .mha (MetaImage) - Modality: CT from a PET/CT acquisition - Units: Hounsfield Units (HU) - No specific slice thickness required — the algorithm resamples to 1 mm isotropic internally before patch extraction

2. PET image (pet-image) - Format: .mha (MetaImage) - Modality: PET from a PET/CT acquisition (typically FDG) - Units: Standardised Uptake Values (SUV) — the image must already be converted to SUV before upload. Raw scanner counts are not accepted. - The image is resampled to 1 mm isotropic internally

3. Nodule locations (points-of-interest) - Format: JSON following the Grand Challenge multiple-points interface - Each point specifies the world (physical) coordinate of one nodule centre, in millimetres, using LPS (Left–Posterior–Superior) convention — the same coordinate system used by SimpleITK and standard DICOM

Example points-of-interest.json:

{
    "name": "Points of interest",
    "type": "Multiple points",
    "points": [
        {
            "name": "Nodule 1",
            "point": [-12.5, 34.0, -150.3],
            "probability": 0.5,
            "time": 0
        }
    ],
    "version": {"major": 1, "minor": 0}
}

Coordinate system: The point field must contain ITK world coordinates (also called physical coordinates) in millimetres. These are the coordinates reported by SimpleITK's TransformIndexToPhysicalPoint, ITK-SNAP's world coordinate display, and stored in DICOM as the Image Position Patient tag. They encode the real-world position of the nodule centre in 3D space, independent of image resolution or field of view.

If you are working in 3D Slicer, note that it displays RAS coordinates — convert to LPS/ITK world coordinates by negating x and y: [-x_RAS, -y_RAS, z_RAS].

Preprocessing pipeline¶

For each nodule, the algorithm: 1. Resamples the CT to 1 mm isotropic spacing (linear interpolation) 2. Resamples the PET to 1 mm isotropic spacing (nearest-neighbour interpolation) 3. Extracts 100 × 100 × 100 mm³ patches centred on the nodule 4. Stacks PET and CT as a two-channel 3D volume and passes it through each of the five ensemble models

Output¶

lung-nodule-malginancy-likelihoods — a JSON object in the same format as the input points-of-interest, with the probability field on each point replaced by the model's calibrated malignancy probability (0 = benign, 1 = malignant).

Interfaces

This algorithm implements all of the following input-output combinations:

Validation and Performance

The model was developed and validated on PET/CT data from Radboudumc.

Uses and Directions

Intended use: This algorithm is intended exclusively for research purposes, such as external validation of results, benchmarking, or exploratory analysis on retrospective PET/CT data. It is not approved or validated for clinical use.

Target population: Adult patients undergoing diagnostic PET/CT with one or more pulmonary nodules requiring malignancy risk assessment.

Appropriate inputs: - FDG PET/CT acquisitions converted to SUV (PET) - Nodules detectable on PET/CT - One prediction is generated per submitted nodule location

Benefits: - Provides a quantitative, reproducible malignancy risk score - Processes multiple nodules per scan in a single submission - Output probabilities are calibrated to reflect true event rates in the development population

Warnings

• PET image must be in SUV units. Submitting raw scanner counts will produce incorrect predictions without any error message.
• Coordinate convention matters. Supplying RAS coordinates (e.g. directly from 3D Slicer without conversion) will result in the wrong image region being sampled.
• Not for clinical use. This model has not been validated or approved for clinical decision-making. It must not be used to guide patient care.
• Out-of-bounds nodules. If a nodule coordinate falls outside or very near the edge of the image, the patch will be zero-padded; predictions for such nodules may be unreliable.

Common Error Messages

The authors have received no reports of common errors. Please carefully check the Warnings.