A radiomics extraction parameterization analysis for machine learning based regression models

• Κύριος συγγραφέας: Λαμπρινάκος, Ιωάννης
• Άλλοι συγγραφείς: Lamprinakos, Ioannis
• Γλώσσα: English
• Έκδοση: 2023
• Θέματα: Radiomics extraction; Neuroendocrine tumors; Machine learning; Artificial intelligence; Peptide receptor radionuclide therapy; Χαρακτηριστικά ακτινομικής; Νευροενδοκρινείς όγκοι; Μηχανική μάθηση; Τεχνητή νοημοσύνη; Ραδιονουκλεϊκή θεραπεία υποδοχών πεπτιδίου

Neuroendocrine Tumors (NETs) are inhomogeneous neoplasms, the most common origin sites of which being the gastrointestinal and pulmonary system. Many treatment methods have been developed to treat this type of cancer, as well as combinations of these are applied. One of these methods is Peptide Receptor Radionuclide Therapy (PRRT) that uses radiation, which is delivered continuously at a decreasing rate to cause retention and/or tumor cell death. Lutathera®, or 177Lu-DOTA-TATE as scientifically named, is a radiolabeled somatostatin analog that is both FDA and EMA approved for PRRT. Radiomics, a non-invasive and quantitative mining medical imaging information tool, are used to extract information from medical images and contribute to diagnosis as well as cancer prognosis, by improving diagnostic accuracy, assisting differential diagnosis between benign and malignant lesions, as well as classifying the risk of disease progression. These quantitative characteristics, which are known as Radiomic Features, can be thus extracted and used for theragnostic reasons. This thesis aims to investigate the ability of the radiomic features that were extracted from images of 68Ga-DOTA-TOC and 177Lu-DOTA-TATE from 21 patients with NETs who underwent therapy with 177Lu-DOTA-TATE to predict the post-therapeutic doses. Four different experiments were conducted and every time parameter values such as Voxel Size and Bin Width changed. Then, in every experiment, image filters named “Wavelets” were once extracted and once omitted. Therefore, 3412 features were extracted from each ROI, including wavelets and 436 features were extracted, excluding wavelets. Moreover, features were selected by following a strict pipeline including Mutual Information regression, efficiency with Pearson and Spearman correlations and, lastly, Backward Elimination Sequential Feature Selection (SFS) and Recursive Feature Elimination (RFE) were used as wrapper methods. Then, a predictive machine learning regression model that included 9 algorithms, both linear and nonlinear, was developed. The performance of this model was judged by taking into account two different metrics, Mean Absolute Error (MAE) and R2. The analysis of our results revealed that when the parameter values (Bin Width, Voxel Size) decreased, more accurate results were acquired. Also, we observed that linear based algorithms seemed to be more efficient than nonlinear ones, with Ridge algorithm being the most prevailing in every experiment for both metrics, and Linear Regression algorithm being the second most prevailing. We also concluded that wavelets-based features were of great importance in our study, since when they were included in our model, better results were yielded.