Research Information System
Thesis Type: Expertise In Medicine
Institution Of The Thesis: Marmara University, School of Medicine, Internal Medical Sciences, Turkey
Approval Date: 2024
Thesis Language: Turkish
Student: Tuğba Nergiz Kıssa Bolat
Supervisor: Fuat Dede
Abstract:
Objective: The selection of suitable patients for TARE in the context of metastatic colorectal cancer is of paramount importance, as it allows for the avoidance of both unnecessary side effects and high costs, which are commonplace in the treatment of other cancers. In order to achieve this, radiomics-based analyses have begun to be incorporated into the prediction of response to TARE treatment, with morphological images such as CT and MRI being employed with great frequency. Furthermore, although there are only a few examples, studies that include FDG PET and metabolic radiomic features have recently begun to appear in the literature. However, the fact that FDG PET was applied in heterogeneous groups comprising different cancer types and that the investigated population was formed without the distinction of Y-90 resin or glass microspheres represents a significant limitation of the few existing studies. In this single-centre study, we sought to investigate the potential of pre-TARE FDG PET radiomics-based modelling to predict the treatment response of the target lesion in patients with colorectal cancer metastatic to the liver who underwent TARE treatment utilising solely glass microspheres.
Methods: A retrospective analysis was conducted on 82 index lesions from 65 patients with metastatic colorectal cancer to the liver who underwent TARE treatment and for whom pre-treatment FDG PET/CT images were available. The lesions were classified into two categories according to PERCIST criteria: treatment-responsive (complete response/partial response) and treatment-refractory (stable disease/progressed disease). The index lesions were manually and blindly segmented by two nuclear medicine physicians using the 3D Slicer programme. Pyradiomics was employed to extract textural and morphological radiomic features. The features demonstrating high inter-observer reproducibility were identified through the use of the Intraclass Correlation Coefficient (ICC). Only those features with ICC values exceeding 0.80 were selected for further analysis. The data were scaled using the StandardScaler function. The issue of class imbalance in the training set was addressed through the use of SMOTE. A machine learning model was constructed using radiomic and clinical features, which were selected using the minimum redundancy maximum relevance (MRMR) algorithm. The performance of the model was evaluated on the training set using the stratified K-fold cross-validation method. The performance of the models was evaluated using a series of metrics, including the area under the curve (AUC), accuracy, sensitivity and specificity.
Results: A total of 82 index lesions were identified in 65 patients (24 females, 41 males, mean age 63.6±11.5 years). Of the lesions, 57.3% demonstrated a response to treatment, while 42.7% did not. A total of 63.4% of cases exhibited evidence of necrosis, while 36.6% did not. No statistically significant differences were observed between the treatment-responsive and non-responsive groups with regard to age, gender, tumour necrosis, volume, 3D Slicer, pre-treatment SULpeak and pre-treatment TLG. The number of radiomic features extracted with pyradiomics was reduced to 56, in accordance with the criterion of ICC>0.80. The number was subsequently reduced to 11 through the application of the MRMR feature selection algorithm. In a clinical context, the MRMR algorithm identified TLG and tumour necrosis as the most relevant variables. The construction of machine learning models based on radiomics-only features and a combination of radiomics and clinical features demonstrated considerable efficacy in predicting the early response to TARE treatment. The Extra Trees model demonstrated the highest performance among the machine learning models created with radiomic features only, followed by the Logistic Regression, Naive Bayes, KNN, SVM and Random Forest models, respectively. In the machine learning models that included clinical features, the Extra Trees, Bagging, Naive Bayes, Random Forest and Adaboost models demonstrated the highest performance, in that order.
Conclusion: The findings of this study indicate that radiomics models based on FDG PET images obtained prior to TARE treatment are an effective means of predicting treatment response in patients with colorectal cancer metastatic to the liver. The high performance of radiomics models and clinical and radiomics models in predicting treatment response indicates that these approaches have significant potential for optimising treatment planning. This approach allows for the optimisation of treatment planning, thereby conferring greater benefits to patients through the implementation of individualised treatment approaches.