ESTRO 2024 - Abstract Book

S4465

Physics - Machine learning models and clinical applications

ESTRO 2024

The details of the multi-omics data and the AI-enriched data included in the RADIANT are listed below. The overlap among different types of data is visualized in Figure 2 .

• Clinical data. The pre-randomisation, in-trial and follow-up records of the trial participants (N = 589). The baseline characteristics, treatment compliance, acute toxicity, treatment response and long-term outcomes were made available from the clinical trial unit. • Radiotherapy data. The radiotherapy data including the planning CT scans, the structure files and the dose distributions were available for 518 patients. All anatomic structures in the planning CT scans were automatically segmented by the Total Segmentator. The dose maps and AI-contoured structures were used to model radiotherapy-induced toxicity. • Pathology data. The digitised H&E-stained WSIs from formalin-fixed paraffin-embedded samples of the pre treatment biopsies (N = 430) and post-treatment resections (N = 417) were available. We applied our AI framework for processing the WSIs. First, the tissue type of each tile (224 pixels ×224 pixels with a pixel size of 0.5 µm) was classified as one of the nine tissue types and the regions of interest (ROIs) were identified according to the tile distribution for further analysis. Second, the five types of nuclei were detected and classified within the ROIs. The proportion and distribution of AI-classified tissues and nuclei were used for studying the tumour microenvironment (TuME). • Genetic and transcriptomic data. The tumour molecular profiling was performed for a sub-cohort randomly selected from the whole trial cohort. The RNA array (N = 300), methylation (N = 124), CNA chromosome segment data (N = 290) and the DNA mutation data (N = 286) from a targeted next-generation-sequencing bespoke panel with 80 colorectal cancer driver genes were collected.

Currently, the RADIANT database is being used for hypothesis-driven post hoc studies, including (1) modelling the post-radiotherapy diarrhoea severity; (2) exploring the impact of the TuME and tumour mutational status and their interaction with the CRT; and (3) predicting the treatment compliance and CRT outcomes using multimodal AI.

Conclusion: