Investigation of radiomics based intra-patient inter-tumor heterogeneity and the impact of tumor subsampling strategies

Abstract

While radiomics analysis has been applied for localized cancer disease, its application to the metastatic setting involves a non-exhaustive lesion subsampling strategy which may sidestep the intrapatient tumoral heterogeneity, hindering the reproducibility and the therapeutic response performance. Our aim was to evaluate if radiomics features can capture intertumoral intrapatient heterogeneity, and the impact of tumor subsampling on the computed heterogeneity. To this end, We delineated and extracted radiomics features of all visible tumors from single acquisition pre-treatment computed tomography of patients with metastatic lung cancer (cohort L) and confirmed our results on a larger cohort of patients with metastatic melanoma (cohort M). To quantify the captured heterogeneity, the absolute coefficient of variation (CV) of each radiomics index was calculated at the patient-level and a sensitivity analysis was performed using only a subset of all extracted features robust to the segmentation step. The extent of information loss by six commonly used tumor sampling strategies was then assessed. A total of 602 lesions were segmented from 43 patients (median age 57, 4.9% female). All robust radiomics indexes exhibited at least 20% of variation with significant heterogeneity both in heavily and oligo metastasized patients, and also at the organ level. None of the segmentation subsampling strategies were able to recover the true tumoral heterogeneity obtained by exhaustive tumor sampling. Image-based inter-tumor intra-patient heterogeneity can be successfully grasped by radiomics analyses. Failing to take into account this kind of heterogeneity will lead to inconsistent predictive algorithms. Guidelines to standardize the tumor sampling step and/or AI-driven tools to alleviate the segmentation effort are required.

Figure 1

Radiomic analysis strategy: from one lesion radiomics to metastatic disease data extraction. While the analysis pipeline for single lesion disease is now described, the metastatic disease pipeline still has some grey areas: first there is no existent guidelines for the sampling strategies of tumors (which one? how many?), and second, there is no consensus on how to aggregate the extracted features for further analysis. Indeed, the number of lesions delineated per patient could vary, and the downstream machine learning pipeline used for single lesion analysis cannot be reused «as-is». Created with BioRender.com.

Figure 2

General view of the data pipeline and the analysis process.

Figure 3

Clustered correlation heatmap of the 27 robust radiomic features. The absolute value of the correlation between features was displayed, for ease of interpretation. 8 clusters were distinguishable, and one radiomic features per cluster were selected for the sensitivity analysis.

Figure 4

Absolute coefficient of variation considering (A) all 107 radiomic features; (B) eight robust and uncorrelated radiomic features (Sensitivity analysis) for cohort L. Radiomics features are subdivided by patient (one color per patient) and the commonly used radiomic features’ classes. All radiomics features exhibited high level of variation for each of the three patients, with shape and glszm features being the most affected. Theses levels of variation were similar with the robust and uncorrelated radiomic features’ set. glcm: Gray Level Co-occurrence Matrix ; gldm: Gray Level Dependence Matrix; glrlm: Gray Level Run Length Matrix; glszm: Gray Level Size Zone; ngtdm: Neighbouring Gray Tone Difference Matrix.

Figure 5

(A) Dissimilarity between lesions’ pairs for patient A. (B) and (C) Volume rendering of the contoured lesions (face + profile). Patient A suffered from bone, lung, pleural peritoneal and soft tissue tumors. Most bone tumors clustered together but showed high dissimilarities with non-bone tumors and even with a small cluster of two bone lesions. There was no obvious clustering of the non-bone lesions (pleural, lung and peritoneal lesions), even when the lesions arised from the same tissue (e.g. pleural lesions).

Figure 6

Recovery curves of maximal tumoral divergence and average tumoral heterogeneity for cohort L. Commonly used subsampling strategies select up to 3–4 lesions, failing to correctly capture intra-patient inter-tumor heterogeneity. Even when the delineation effort is scaled up to 15 lesions, none of the two heterogeneity indexes are fully recovered.