Medical imaging is a powerful tool for clinical practice allowing in-vivo insight into a patient’s disease state. Many modalities exist, allowing for the collection of diverse information about the underlying tissue structure and/or function. Traditionally, medical professionals use visual assessment of scans to search for disease, assess relevant disease predictors and propose clinical intervention steps. However, the imaging data contain potentially useful information beyond visual assessment by trained professional. To better use the full depth of information contained in the image sets, quantitative imaging characteristics (QICs), can be extracted using mathematical and statistical operations on regions or volumes of interests. The process of using QICs is a pipeline typically involving image acquisition, segmentation, feature extraction, set qualification and analysis of informatics. These descriptors can be integrated into classification methods focused on differentiating between disease states. Lung cancer, a leading cause of death worldwide, is a clear application for advanced in-vivo imaging based classification methods.
We hypothesize that QICs extracted from spatially-linked and size-standardized regions of surrounding lung tissue can improve risk assessment quality over features extracted from only the lung tumor, or nodule, regions. We require a robust and flexible pipeline for the extraction and selection of disease QICs in computed tomography (CT). This includes creating an optimized method for feature extraction, reduction, selection, and predictive analysis which could be applied to a multitude of disease imaging problems. This thesis expanded a developmental pipeline for machine learning using a large multicenter controlled CT dataset of lung nodules to extract CT QICs from the nodule, surrounding parenchyma, and greater lung volume and explore CT feature interconnectivity. Furthermore, it created a validated pipeline that is more computationally and time efficient and with stability of performance. The modularity of the optimized pipeline facilitates broader application of the tool for applications beyond CT identified pulmonary nodules.
We have developed a flexible and robust pipeline for the extraction and selection of Quantitative Imaging Characteristics for Risk Assessment from the Tumor and its Environment (QIC-RATE). The results presented in this thesis support our hypothesis, showing that classification of lung and breast tumors is improved through inclusion of peritumoral signal. Optimal performance in the lung application achieved with the QIC-RATE tool incorporating 75% of the nodule diameter equivalent in perinodular parenchyma with a development performance of 100% accuracy. The stability of performance was reflected in the maintained high accuracy (98%) in the independent validation dataset of 100 CT from a separate institution. In the breast QIC-RATE application, optimal performance was achieved using 25% of the tumor diameter in breast tissue with 90% accuracy in development, 82% in validation. We address the need for more complex assessments of medically imaged tumors through the QIC-RATE pipeline; a modular, scalable, transferrable pipeline for extracting, reducing and selecting, and training a classification tool based on QICs. Altogether, this research has resulted in a risk assessment methodology that is validated, stable, high performing, adaptable, and transparent.