Towards the use of multiple ROIs for radiomics-based survival modelling: Finding a strategy of aggregating lesions

Abstract

Background:

Radiomic features, derived from a region of interest (ROI) in medical images, are valuable as prognostic factors. Selecting an appropriate ROI is critical, and many recent studies have focused on leveraging multiple ROIs by segmenting analogous regions across patients — such as the primary tumour and peritumoral area or subregions of the tumour. These can be straightforwardly incorporated into models as additional features. However, a more complex scenario arises, for example, in a regionally disseminated disease, when multiple distinct lesions are present.

Aim:

This study aims to evaluate the feasibility of integrating radiomic data from multiple lesions into survival models. We explore strategies for incorporating these ROIs and hypothesize that including all available lesions can improve model performance.

Methods:

While each lesion produces a feature vector, the desired result is a unified prediction. We propose methods to aggregate either the feature vectors to form a representative one or the modelling results to compute a consolidated risk score. As a proof of concept, we apply these strategies to predict distant metastasis risk in a cohort of 115 non-small cell lung cancer patients, 60% of whom exhibit regionally advanced disease. Two feature sets (radiomics extracted from PET and PET interpolated to CT resolution) are tested across various survival models using a Monte Carlo Cross-Validation framework.

Results:

Across both feature sets, incorporating all available lesions — rather than limiting analysis to the primary tumour — consistently improved the c-index, irrespective of the survival model used. The highest c-Index obtained by a primary tumour-only model was 0.611 for the PET dataset and 0.614 for the PET_CT dataset, while by using all lesions we were able to achieve c-Indices of 0.632 and 0.634.

Conclusion:

Lesions beyond the primary tumour carry information that should be utilized in radiomics-based models to enhance predictive ability.