Radiological tumor classification across imaging modality and histology.

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Radiomics uses quantitative features from radiological scans for imaging biomarkers.
• Current radiomic signatures lack reproducibility and generalizability due to modality and histology dependence.
• Novel features are needed for robust tumor characterization across diverse imaging protocols.

Purpose of the Study:

• To develop novel radiological features for compatible and reproducible tumor characterization.
• To identify and validate unifying imaging subtypes across multiple cancer types and imaging modalities.
• To assess the clinical utility of these subtypes in predicting prognosis and treatment response.

Main Methods:

• Extraction of novel quantitative radiological features designed for cross-modality and cross-tissue compatibility.
• International multi-institution study involving 1,682 patients across three malignancies and two imaging modalities.
• Deep learning for automated tumor segmentation and subtype identification.

Main Results:

• Discovery and validation of four unifying imaging tumor subtypes.
• These subtypes exhibit distinct molecular characteristics and prognoses.
• In advanced lung cancer treated with immunotherapy, one subtype showed improved survival and increased tumor-infiltrating lymphocytes.

Conclusions:

• The proposed radiomic features enable systematic characterization of tumor morphology and spatial heterogeneity.
• The identified unifying tumor subtypes offer potential for improved prognosis and treatment response prediction.
• Deep learning-enabled implementation facilitates practical application in precision medicine.