Development and application of AI tools to improve diagnostic and prognostic capabilities of medical imaging data
Date issued
Authors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
License
Abstract
Advances in personalized cancer treatment have been significantly motivated by the integration of advanced computational techniques with medical imaging modalities. These techniques allow for the extraction of quantitative radiomics information, which is then analyzed and correlated with genomic biomarkers, clinical features, and other biological indicators through machine learning (ML) models. This computational approach allows for a better tracking of cancer behavior and treatment response, offering a powerful tool for diagnostic, therapeutic planning and prognosis purposes.
In this thesis, ML models were developed and validated to assess the predictive power of quantitative radiomics and radiogenomic features, extracted from magnetic resonance imaging (MRI) and computed tomography (CT) scans, as well as clinical and semantics information. These models were analyzed in different oncological contexts, using computational methods to process all features, including batch harmonization, outlier detection, normalization, feature selection via redundancy reduction and relevance optimization and class imbalance correction. Classifiers employed included support vector machine (SVM), random forest (RF), least absolute shrinkage and selection operator (LASSO), logistic regression (LR) and multilayer perceptron (MLP) classifiers.
The first study of this thesis focused on the differentiation between atypical lipomatous tumors (ALTs) and lipomas via the detection of the mouse double minute 2 (MDM2) gene biomarker. A LASSO classifier performed best with an area under the receiver-operator characteristic (AUROC) of 0.88. The second study investigated the complete pain response in painful spinal bone metastasis patients after palliative radiotherapy (RT) treatment. A clinical LASSO classifier achieved an AUROC of 0.80. The third study explored common acute side effects on breast cancer patients treated with RT. A LASSO classifier outperformed all other models when predicting the appearance of moist cells epitheliolysis as a surrogate for skin inflammation with an AUROC of 0.74. The fourth study monitored neoadjuvant chemotherapy treatment response to Ewing sarcoma patients via the histological response assessment after surgery. A LR trained on the relative delta of radiomics features achieved an AUROC of 0.62, outperforming the best model trained on information from radiology readings (LR; AUROC of 0.58).
In conclusion, this thesis provides insight into the value of artificial intelligence (AI) and radiomics to address key challenges in oncology by supporting clinical decisions regarding distinguishing tumor types, predicting treatment responses, toxicity and disease evolution. Radiomics features were most effective when differentiating visually similar tumors and as a relative delta of change before and after neoadjuvant chemotherapy treatment. Clinical features have also shown predictive power in other cases of treatment response prediction, while providing useful support and baseline information overall.