Utilising Quantitative Imaging And Radiomics To Unravel Intra-tumoural Heterogeneity In Soft Tissue Sarcoma

Abstract

Soft tissue sarcomas (STS) are rare mesenchymal tumours that exhibit extensive biological and clinical heterogeneity. Our understanding of this heterogeneity is incomplete and has continued to be a barrier to clinical advancements, and as such, outcomes for STS patients remain poor. Across oncology, radiology has transformed into a data-driven specialty, where advanced imaging techniques and computational science such as quantitative imaging and radiomics have the potential to provide in-depth information about tumour biology. If clinically translated, these tools could provide tumour characterisation by probing biology in a non-invasive and global manner revolutionising STS diagnostics and prognostics. However, progression of these tools has been slow as we continue to lack an understanding of the biology underlying imaging outputs. Herein, my project aimed to utilise imaging to approach biological heterogeneity in retroperitoneal sarcoma (RPS) in two ways: direct radiological-pathological correlation and large-scale CT-based radiomics. Using a retrospective cohort of 21 patients with RPS, immunohistochemistry of multi-regional tissue matched to regions of interest on quantitative MRI (qMRI) were analysed. Specifically key immune cells were characterised and correlation analysis with qMRI and histopathological data was carried out. Expansion to patient level data was conducted to include clinical data for correlation. Key findings were intra-tumoural heterogeneity of immune cells, the correlation of MRI enhancing fraction with T cells and the correlation of Ki67 proliferation index with T cells and CD68+ cells at both block and patient levels. To interrogate further the capability of unravelling heterogeneity using imaging, radiomic and deep learning (DL-) radiomic data available for the same cohort was analysed. Correlation analysis revealed 32 correlations between immune and histopathological data with DL-radiomics whilst 5 correlations were found between histopathological and radiomic data, with absence of any correlations between immune and radiomic data. In addition, clustering of DL-radiomics data revealed a persistent cluster inclusive of all recurrence cases alluding to a possible DL-radiomic signature for RPS disease recurrence. Using large-scale CT-based radiomics, we developed and tested a classification model for the two most common RPS subtypes, leiomyosarcoma (LMS) and liposarcoma (LPS) and for histological grade. Our model utilised a novel feature selection pipeline, sub-segmentation to provide sub-region volumes, repeatability analysis and intense cross-validation to increase the explainability and stability of our final models. Selecting the best performing models for prediction of subtype and for grade, we successfully independently validated our models using international, multi-centre trial data. Final area under the receiver operator curves were 0.928 for subtype and 0.882 for low versus intermediate/high grade RPS. Compared with standard of care radiological and histopathological assessment of subtype and grade, our models outperformed with regards to accuracy: 0.843 (radiomics) versus 0.65 (clinical) for subtype and 0.823 (radiomics) versus 0.44 (clinical) for low versus intermediate/high grade. These powerful and validated radiomics models, if prospectively validated could revolutionise the upfront characterisation of RPS tumours, offering a complimentary tool for diagnosis and risk stratification. Overall, my project has provided an exciting foundation of using qMRI and radiomics to provide a virtual biopsy of underlying biology of RPS tumours. It qualifies as a rich resource to guide future work to ultimately encourage the clinical translation of cutting-edge tools that may be utilised in diagnosis, risk stratification, treatment planning and monitoring and prognosis of STS.