Disease trajectories describe the health of patients as they are diagnosed, treated, and observed. Along this path, physicians are repeatedly expected to integrate numerous pieces of evidence and explain the possible outcomes of the disease, such as survival. This expectation becomes increasingly challenging with the proliferation of patient data collected. The goal of this research is to evaluate the use of machine learning and related techniques to integrate complementary patient data to better characterize disease trajectories. This dissertation provides approaches (1) to map cellular-level information to diagnostic images, (2) to identify temporal patterns that can estimate the probability of decline at each patient encounter, and (3) to interpret the trained models to explain the underlying patterns of disease and subset ones that affect survival. The approaches are demonstrated on patients with glioblastoma, a brain tumor with poor prognosis.

This research shows how radiogenomic maps can be created using deep neural networks to improve the representation of gene expression. Methods developed to interpret neural networks were systematically evaluated and found transcriptomic patterns of imaging traits that were both new and consistent with prior works. Furthermore, longitudinal changes were mined from patients' radiology and oncology visits. The temporal patterns were used to predict residual survival at each visit and were consistent with physicians' understanding of disease progression. Together, these approaches demonstrate how patterns integrated from molecular, medical imaging, and other clinical data can provide prognostic insights throughout a patient's disease trajectory.

Objective

Method

Results

Conclusions

Objective

Method

Results

Conclusions

Low-dose computed tomography (CT) screening has been widely used to detect and diagnose early stage lung cancer. Clinical trials have shown that low-dose CT reduced lung cancer mortality by 20% relative to plain chest radiography; however, challenges exist in current low-dose CT screening programs including high over-diagnosis rates, high cost and increased radiation exposure. This dissertation attempts to overcome these challenges by developing machine and deep learning models for automated lung cancer diagnosis and disease progression estimation. A novel lung segmentation approach was first developed using a bidirectional chain code method and machine learning framework. This method is designed to include the lung nodules attached to lung wall while minimizing over-segmentation error. Second, a hybrid ensemble convolutional neural network has been developed to classify lung nodule vs. non-nodule objects. The ensemble model combines the VGG, residual and densely connected module designs to improve the model classification robustness for external datasets collected with different acquisition parameters. Third, a hierarchical semantic convolutional neural network (HSCNN) has been described to classify lung nodule malignancy. Semantic characteristic features, predicted in parallel with the malignancy for each nodule, enable the interpretation of the model and improvement of malignancy prediction. Finally, a Bayesian framework combined with a continuous-time Markov model was developed to estimate the multi-state disease progression of lung cancer. The resulting model estimates individual lung cancer state transition information, providing the basis for personalized screening recommendations. Extensive experiments and results have proved the effectiveness of these methods paving the way to optimize and improve the effectiveness of existing low-dose CT screening programs.

Lung cancer is the leading cause of cancer-related death in both sexes. Large, randomized clinical trials have demonstrated that low-dose computed tomography screening reduces mortality from lung cancer, as opposed to chest X-ray or no screening, when participants adhere to follow-up recommendations. However, low adherence rates in post-trial clinical lung cancer screening (LCS) programs have been reported across the United States (US). Low adherence to LCS in real-world clinical practice diminishes the mortality benefit of annual screening derived from clinical trials. Thus far, limited studies have examined factors affecting the patient’s decision (not) to adhere to screening guidelines. This dissertation examines the factors that may predict patient non-adherence to LCS recommendations. First, we performed a systematic review and meta-analysis of 24 studies published between 2014 and 2020 that mentioned adherence to Lung-RADS recommendations, identifying factors contributing to adherence rates. Second, using the Carter-Harris conceptual model, which enumerates psychosocial variables (e.g., smoking stigma, cancer fear, cancer fatalism) related to LCS participation and adherence, we examined these variables’ availability and completeness in our medical records. Next, using a subset of variables where data were consistently available, we identified factors of non-adherence over multiple screening time points using logistic regression and mixed effects models. Lastly, we used statistical and machine learning-based methods to examine how well we could predict patient non-adherence using longitudinal data across three screening time points. This dissertation advances our understanding of factors contributing to patient non-adherence to LCS recommendations. It provides a basis for identifying patient groups that could benefit from individualized interventions to improve LCS adherence.

Objective

Method

Results

Conclusions

Computed tomography (CT) plays an integral role in diagnosing and screening various types of diseases. A growing number of machine learning (ML) models have been developed for prediction and classification that utilize derived quantitative image features, thanks in part to the availability of large CT datasets and advances in medical image analysis. Researchers have classified disease severity using quantitative image features such as hand-crafted radiomic and deep features. Despite reporting high classification performance, these models typically do not generalize well. Variations in the appearance of CT scans caused by differences in acquisition and reconstruction parameters adversely impact the reproducibility of quantitative image features and the performance of machine learning algorithms. As a result, few ML algorithms have been used in clinical settings. Mitigating the effects of varying CT acquisition and reconstruction parameters is a challenging inverse problem. Recent advances in deep learning have demonstrated that image translation and denoising models can achieve high per-pixel similarity metrics when compared to a target image. The purpose of this dissertation is to develop and evaluate two conditional generative models that mitigate the effects of working with CT scans acquired and reconstructed with a variety of parameters. The overarching hypothesis is that improved image quality results in better consistency in nodule detection. In essence, these models attempt to learn the underlying conditional distribution on the normalized images (high-quality) given the un-normalized (low-quality) images. First, I propose a novel CT image normalization method based on a 3D conditional generative adversarial network (GAN) that utilizes a spectral-normalization algorithm. My model provides an end-to-end solution for normalizing scans acquired using different doses, slice thicknesses, and reconstruction kernels. This study demonstrates that the GAN is capable of mitigating the variability in image quality, quantitative image features, and lung nodule detection using an automated Computer-Aided-Detection (CAD) algorithm. I show that GAN improved perceptual similarity by 22%, and resulted in a 16% increase in features with a good level of agreement based on concordance correlation coefficient analysis. As a result, the performance of the existing nodule detection model was up to 75% more consistent with the reference scan. Second, I explore the use of a conditional normalizing flow-based model to incorporate uncertainty information during image translation. The model is capable of learning the explicit conditional density and generating several plausible image outputs, providing a means to reduce the distortions introduced by existing methods. I show that the normalizing flow method achieves a 6% improvement in perpetual quality compared to the state-of-the-art GAN-based method and the resulted agreement level of the detection task is improved by 13%. This dissertation compares these two generative approaches, identifying their strengths and limitations when normalizing heterogeneous CT images and mitigating the effect of different acquisition and reconstruction parameters on downstream clinical tasks.