CIRDataset

CIRDataset provides a curated dataset and associated deep learning models for clinically-interpretable radiomic analysis of lung nodule spiculations and lobulations to support malignancy prediction.

Key Features:

• Extensive Annotations: 956 annotations of spiculations and lobulations on segmented lung nodules with radiologist QA/QC to ensure clinical accuracy.
• Comprehensive Data Sources: Aggregated cases from LIDC-IDRI (N=883) and LUNGx (N=73).
• Multi-class Voxel2Mesh Model: An end-to-end deep learning model based on a multi-class Voxel2Mesh extension that segments nodules while preserving spike geometry, classifies spikes into sharp/spiculation and curved/lobulation categories, and predicts malignancy.
• Integration of Clinically-Reported Features: Incorporation of clinically-relevant features (spiculations and lobulations) to enable benchmarking and validation of malignancy prediction algorithms.

Scientific Applications:

• Lung Nodule Malignancy Prediction: Development and evaluation of algorithms that predict malignancy using annotated spiculation and lobulation features.
• Radiomics Interpretability: Research to improve interpretability of AI-driven predictions by linking radiomic features to clinical criteria such as Lung-RADS.
• Algorithm Benchmarking and Validation: Quantitative comparison and validation of segmentation, spike classification, and malignancy prediction methods against radiologist-annotated ground truth.

Methodology:

Radiologists performed meticulous annotations and QA/QC of spiculations and lobulations which were integrated into an end-to-end deep learning framework using a multi-class Voxel2Mesh extension for segmentation, spike classification, and malignancy prediction.