ATTRACT - ArTificial inTelligence-based RAdiogenomics in Colon Tumors

The ATTRACT trial consist of a retrospective and a prospective part. In the retrospective part of the trial, radiology, pathology, and genomics data of 300 patients with stage II-III colon cancer will be used to identify the genetic and radiomics features of colon tumors and the clinical endpoints as the outcomes of the predictive model.

Tumors will be manually segmented on CT images and used for the AI (artificial intelligence)-model generation. Pathological annotations will be associated to the corresponding anonymised profiles. Immunohistochemistry will be used to classify the samples in the 4 molecular subtypes according.RNA-seq profiles will be also generated from tissue samples through targeted transcriptomics using custom NGS (next-generation sequencing) panels specifically designed to evaluate gene expression and assess Tumor Mutational Burden (TMB). Raw data will be processed and modelled using Topological Pathway Analysis to stratify patients according to the relevant molecular features and define molecular annotations that will be used to train the model for the identification of specific clinically relevant groups. Raw data together with radiological data will be used to generate and train the AI-models for the automated segmentation and the extraction of the radiogenomics signature.

Radiomics features will be extracted from manually segmented tumors. Standard PyRadiomics tools as well custom-made tools will be used. Feature robustness will be guaranteed by selecting only those with high inter-observer statistical correlation. Two families of AI models will be generated, one family dedicated to segmentation, and the other dedicated to radiogenomics-based phenotyping according to the clinical, molecular biology and pathological data available. The two families will be fused for the creation of the ATTRACT AI-model. For the generation of these models, specific convolutional neural network (CNN) architectures based on deep learning (DL) and Artificial Intelligence like UNet and MaskNet will be applied. The training will be performed using the manual ROI (region of interest) segmentations as ground truth. For the generation of radiogenomics analysis models, radiomics and genomic features will be combined using different multivariate algorithms. The classificatory will be trained to recognise the cancer subtypes and clinical endpoints.

In the prospective part of the trial, patients with stage II-III colon cancer will be recruited and will undergo a preoperative contrast-enhanced CT examination. The recruitment rate will be 70 patients per year, for a total of 210 patients. After pre-operative CT, surgery will be performed according to international standard protocols. Eventually adjuvant therapy will be considered following current guidelines. Pathological sample of the prospective enrollment will be analyzed. First, with RNA-seq data, TMB (of coding genes) and clinical data, patients will be clustered by making use of two different techniques Markov Cluster algorithm (MCL) and t-SNE (t-distributed stochastic neighbor embedding), Multi-Layer Network clustering. Patients will be represented as node of a network, edges between nodes will be weighted and thresholded according to the Jaccard Similarity. The similarity will be computed on top of Gene Expression, TMB, and perturbation information coming from Topological Pathway Analysis. Results of clustering will be matched with those coming from Immunohistochemistry. Clinical follow-up data (i.e. outcome of the therapy etc...) will be, once available, also plugged into the workflow to enforce the learning. Extracted knowledge will be used to annotate the dataset used to train and validate the radiomics classification. Gross specimen will be analyzed in order to extract different transcriptomics molecular subtypes (CMS1, CMS2, CMS3, CMS4) in accordance to the Colorectal Cancer (CRC) Subtyping Consortium (CRCSC) assessing the presence or absence of core subtype patterns among existing gene expression-based CRC subtyping algorithms. The accordance between pathological molecular profile and ctDNA analysis during protocol will be related to radiomics classification in order to provide a new whole-diagnostic model of approach in CRC treatment and surveillance.

Prospective data will be used to validate the AI models. For the segmentation models, the Dice Coefficient will be used as an indicator to measure the degree of overlap between the automated and the expert segmentation. For the radiogenomics model, performances will be evaluated using accuracy, integral under the receiver operator curve (ROC-AUC) and clinical decision curve. The investigators will also take into consideration, in order to select the best AI models, the response to the variation of the input characteristics and will produce saliency maps where the features of the input image that mostly contributed to the classification are highlighted.