ESTRO 2024 - Abstract Book

S4464

Physics - Machine learning models and clinical applications

ESTRO 2024

Material/Methods:

(1) Data collection and linkage

ARISTOTLE is a phase III clinical trial investigating whether the addition of a second drug (irinotecan) to the standard neoadjuvant treatment of long-course capecitabine and radiotherapy improves outcomes for MRI-defined LARC patients. The anonymised trial ID was used to link clinical data from the UCL Clinical Trial Centre, the radiotherapy & imaging data from the Radiotherapy Trials Quality Assurance group, the pathology data from the University of Leeds and the genomic data from the Stratification in Colorectal Cancer (S: CORT) Consortium, to establish the RADIANT research database.

(2) Multi-task AI processing

To enrich the information for predictive modelling, multi-task AI models were used for processing the image data. The data collection and generation workflow are displayed in Figure 1 . We applied the Total Segmentator [2], an open-source AI model to contour the anatomical structures automatically on the radiotherapy planning CT. An AI pathology framework was designed for the analysis of the digitised Haematoxylin and Eosin (H&E)-stained whole slide images (WSIs). The pathology framework consists of two AI models trained for tissue classification and nuclei detection using open-source datasets. The tissue classifier was trained on the NCT-CRC-HE-100K [3] dataset to identify nine tissue types, including adipose, background, debris, lymphocytes, mucin, smooth muscle, normal colon mucosa, cancer-associated stroma and colorectal adenocarcinoma epithelium. The nuclei detector was trained on the Lizard dataset [4] to detect and classify five types of nuclei, including epithelial nuclei (neoplastic/ non-neoplastic epithelial nuclei), inflammatory nuclei (lymphocyte, plasma, neutrophil and eosinophil) and connective tissue (fibroblasts, muscle and endothelial cells).

Results: