Photon-counting Detector CT Spectral Reconstructions for Radiomics-based Liver Lesion Classification: A Multicenter...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate noninvasive classification of hepatic lesions is challenging with conventional CT.
• Photon Counting Detector CT (PCD-CT) offers spectral imaging for enhanced tissue characterization.
• Radiomics and machine learning show promise for improving diagnostic accuracy.

Purpose of the Study:

• Evaluate radiomics-based machine learning for differentiating benign and malignant liver lesions.
• Utilize multispectral images from contrast-enhanced PCD-CT.
• Assess the impact of different spectral datasets on classification performance.

Main Methods:

• Multicenter study using PCD-CT data from 378 patients with hepatic lesions.
• Automated lesion segmentation using a nnU-Net model.
• Radiomic feature extraction from 9 spectral datasets (40-180 keV, VNC, IDM).
• Patient-based classification using Random Forest models.

Main Results:

• Random Forest model trained on VNC achieved the highest performance (AUC: 0.899, Accuracy: 83.5%).
• Models tested on higher energy levels (110, 140, 180 keV) and VNC showed the most stable performance (median AUC ≥ 0.83).
• Lower energy levels (40, 50 keV) and Iodine Density Maps (IDM) yielded lower and more variable results.

Conclusions:

• Radiomics-based machine learning reliably differentiates benign and malignant hepatic lesions on PCD-CT.
• Higher energy reconstructions and VNC images are crucial for stable classification performance.
• Integrating these findings into clinical workflows can aid noninvasive risk stratification and reduce invasive procedures.