Prediction of Subsolid Pulmonary Nodule Evolution from Baseline CT Using Temporal Imaging Models

Code for tracking subsolid lung nodules across serial CT examinations and predicting their growth using 3D temporal models.

---

Pipeline

Step	Script	Purpose
1	preprocessing/nifti.py	Convert source images and nodule masks to NIfTI
2	preprocessing/lung_masks.py	Generate binary lung-field masks
3	preprocessing/match.py	Deformable registration + cross-timepoint nodule matching
4	preprocessing/growth.py	Compute VDT and growth labels for all nodule pairs
5	preprocessing/precompute.py	Extract 64x64x64 crops for model input
6	models/discriminative.py	Train 3D growth classifier
6	models/generative.py	Train 3D future mask predictor

---

Setup

conda env create -f environment.yml
conda activate nodule-growth

Edit paths.yaml to set data_root and verify all entries point to valid locations on your system. All scripts read paths from this file exclusively.

---

Input Data

The pipeline is agnostic to the upstream detection/segmentation algorithm. Any system that produces the following can be used as input:

Scan Metadata (scan_metadata_csv in paths.yaml)

A CSV with one row per CT scan. Required columns:

Column	Description
PatientID	Unique patient identifier
ScanNumber	Scan index within patient
StudyDate	Acquisition date
Series	Series identifier
SeriesInstanceUID	DICOM SeriesInstanceUID
imagePath	Path to source CT image (DICOM or MHD)
maskPath	Path to nodule segmentation mask

Nodule Annotations (nodule_annotations_csv)

Per-nodule measurements. Required columns:

Column	Description
FindingUID	Unique nodule identifier
PatientID	Patient identifier
SeriesInstanceUID	Which scan this measurement belongs to
StudyDate	Date of measurement
Long	Long-axis diameter (mm)
Short	Short-axis diameter (mm)

Nodule Segmentations (nodule_segmentations_csv)

Per-nodule segmentation metadata. Required columns:

Column	Description
PatientID	Patient identifier
SeriesInstanceUID	Scan identifier
mask_label	Integer label of this nodule within the mask volume
FindingUID	Links to annotation table
NoduleTypeStr	Nodule type (e.g., "ground glass", "Partial solid")
segmentation	JSON containing sitk_volume_mm3

Registration Patient List (registration_pids_csv)

A single-column CSV (PatientID) listing which patients to process.

Patient Splits (train/val/test)

Three single-column CSVs (PatientID) defining patient-level data splits. Used by the models to prevent data leakage.

---

Models

Discriminative — Binary Growth Classifier

3D ResNet predicting whether a nodule is growing (>= 1.5 mm diameter increase on any axis). Takes a 2-channel 64^3 input (CT + mask) conditioned on elapsed time. Trained with class-weight balancing and early stopping on validation AUC.

Generative — Future Mask Predictor

3D U-Net predicting the nodule segmentation mask at a future timepoint. Takes a 3-channel 64^3 input (CT + mask + time map) with learned time embedding at the bottleneck. Trained with a combined Dice + BCE + volume loss, early stopping on validation Dice.

---

Growth Classification

Criterion	Growing	Stable	Shrinking
Volume change	> 25% increase	+/- 25%	> 25% decrease
Diameter change	>= 1.5 mm increase (any axis)	< 1.5 mm	>= 1.5 mm decrease
VDT	< 400 days	>= 400 days	—

Biologically implausible VDT (< 20 or > 5000 days) is excluded.

---

License

MIT License.