579P - A clinico-imaging predictive artificial intelligence model of relapse in colon cancer using baseline CT scans

Background

Up to 35% of patients with localized colon cancer (CC) have a relapse, even after receiving optimal initial treatment. New prognostic biomarkers are required to more accurately identify patients who benefit from receiving adjuvant treatment and more personalized care. We propose an artificial intelligence (AI) model combining clinical data with fractals, radiomics and deep features extracted from baseline computed tomography (CT) exams for the prediction of CC relapse.

Methods

This multi-center retrospective observational study includes real-world clinical data from CC patients and CT examinations from 2015 to 2021. Manual segmentation of the tumor was performed by an experienced abdominal radiologist. A total of 2158 radiomics, deep features and fractal dimension features were extracted using Quibim Precision® software (Quibim, Valencia, ES). Harmonization techniques were applied to address image quality variability. ResNet152V2 deep convolutional network was applied as a transfer learning framework for deep features extraction and logistic regression model to estimate probability of relapse. 80% of the cases (136 studies) were used to train and validate the model following a nested cross-validation strategy to handle overfitting problem while 20% (34) were used for independent test set to analyze model generalization.

Results

Baseline CT exams from 170 localized CC patients (69 relapsed and 101 non-relapsed patients) with stages I (4%), II (20%) and III (76%) were included. Ten features were finally selected (2 radiomics, 8 deep features) by mRMR algorithm, providing an increased area under the curve (AUC) of 0.70 (95% CI: 0.52-0.88) in the independent set higher than that obtained with the model trained with clinical variables alone (0.50, 95% CI: 0.35-0.65). Using the DeLong test to probe the difference between the AUC of models, it can be conclude that model including radiomic information has a statistically different AUC from model with only clinical features with p-value < 0.05 (p = 0.023).

Conclusions

This AI approach, by identifying a panel of imaging biomarkers and in combination with clinicopathological features, will help improve clinical decision-making and patient outcomes using standard-of-care data.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

INCLIVA.

Funding

Has not received any funding.

Disclosure

J. Vidal Barrull: Other, Personal, Speaker, Consultant, Advisor: Merck Serono, Amgen, Pierre Fabre; Other, Personal, Advisory Role: Bristol Myers Squibb. C. Montagut Viladot: Other, Personal, Advisory Role: Amgen, Roche, Sanofi, Merck, Lilly, Guardant, Biocartis. A. Cervantes: Other, Institutional, Advisory Role: Merck Serono, Amgen, Roche, Transgene, Foundation Medicine; Other, Personal, Other: Cancer treatment Reviews, Annals of Oncology, ESMO Open; Other, Institutional, Research Funding: Novartis, BaiGene, FibroGen, Astellas Pharma, MedImmune, Amgen, Actuate Therapeutics, Adaptimmune, AstraZeneca Spain, Amcure, Bayer, Bristol Myers Squibb, Lilly, Genentech, Merck Serono, Natera, MSD, Servier, Sierra Oncology, Takeda. N. Tarazona Llavero: Other, Personal, Advisory Role: Guardant Health; Other, Personal, Speaker’s Bureau: Merck, Amgen, Servier, Pfizer; Other, Personal, Other: Merck, Amgen, Roche. All other authors have declared no conflicts of interest.