Radiomics and machine learning for renal tumor subtype assessment using multiphase computed tomography in a...

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Accurate histological subtyping of renal tumors is crucial for treatment planning.
• Distinguishing between renal tumor subtypes using imaging alone can be challenging.

Purpose of the Study:

• To evaluate the efficacy of radiomic features and machine learning (ML) in differentiating renal tumor histological subtypes.
• To assess the diagnostic performance of ML models based on multiphase computed tomography (CT) data.

Main Methods:

• Retrospective analysis of CT scans from 297 patients with renal tumors.
• Extraction of radiomic features from arterial and venous phase CT scans.
• Development and validation of an extreme gradient boosting (XGB) ML algorithm for subtype classification.

Main Results:

• The XGB model achieved an area under the receiver operating characteristic curve (AUC) of 0.81 for venous phase and 0.8 for combined phases in the training cohort.
• Independent testing showed AUCs of 0.75 for venous and 0.75 for combined phases.
• Angiomyolipomas (AMLs) were identified with high accuracy (AUC 0.9-0.94), while oncocytomas showed the lowest accuracy (AUC 0.57-0.69).

Conclusions:

• Radiomic feature analysis using ML can reliably distinguish renal tumor subtypes on routine CT scans.
• Arterial phase CT radiomic features did not significantly improve subtype identification accuracy compared to venous or combined phases.
• Oncocytomas represent the most difficult subtype to differentiate using this radiomic approach.