Task-Targeted Pre- and Post-acquisition Methodologies for Optimal Conditioning and Interpretation of Medical Images Using Virtual Imaging Trial

The use of medical imaging plays a vital role in diagnosing and monitoring diseases, significantly enabling healthcare providers to provide better patient care. Medical images have much wider significance than just a single clinical evaluation; they are crucial components of developing data-driven algorithms and conducting studies that involve multiple healthcare centers. These algorithms and research studies depend heavily on medical image data consistency and accuracy, which ultimately impact treatment precision and diagnostic accuracy.

The ever-evolving landscape of medical imaging has led to diverse imaging systems and modalities over the past decades. While this diversity offers a wealth of options for generating medical images, it also poses challenges. Due to this diversity, determining the most suitable protocols and parameters for achieving optimal image quality for specific clinical tasks has become a complex task. Consequently, inconsistencies in protocol selection can result in variable quantification outcomes.

The variety of scanning systems and protocols for capturing patient images has made establishing a multidimensional correlation between protocol selection and image quality challenging. Consequently, finding the most appropriate protocols to maximize image quality for specific tasks has become a more complex endeavor. Furthermore, when dealing with previously acquired medical images, the available data is typically limited to the image itself and associated meta-data. Therefore, understanding the governing physics of the acquisition is vital for tailoring post-processing tools to preserve diagnostic information and maintain consistency in quantification.

This proposal sets out to develop patient-centric, data-driven approaches customized for specific diagnostic tasks. These methodologies aim to enhance the consistency and accuracy of medical images, guided by two key aims: - Optimizing the protocol selection: Developing and evaluating optimization-based methodologies for selecting protocols customized for distinct tasks. The emphasis here is on refining medical image acquisition protocols to attain the highest levels of consistency and quality. - Harmonizing the images: Design and validate Deep Neural Network (DNN) architectures focused on post-processing of the images to reduce quantification variability. These DNN-based solutions enhance quantification consistency while preserving essential diagnostic information and the clinical appearance of medical images for visual inspection and interpretation by radiologists.

However, clinical data scarcity poses significant challenges to algorithms designed for medical applications. The limited availability of data hampers the development of effective algorithms. The absence of precise ground truth information restricts medical evaluations to variability assessment, leaving biases and accuracy largely unexplored.

To address the latter problem, we utilized VIT. By leveraging VIT, one can access extensive datasets encompassing a wide range of variations while providing accurate ground truth data. This approach circumvents patient health and privacy concerns, as the data is virtual rather than derived from actual patients. Emphasis has been placed on the chest area, which harbors vital organs such as the heart and lungs, while targeting diseases like coronary stenosis, lung cancer, and chronic obstructive pulmonary disease (COPD), all of which are globally recognized as some of the most lethal diseases.

To optimize the selection guidelines for protocols, correlation models were calculated to capture the governing physics of the acquisitions, allowing us to investigate how changes in protocols may affect the selection of other parameters. Polynomial approximations have been introduced in place of the resource-intensive and computationally demanding algorithms used to assess measurement quality based on sampling. Subsequently, a framework for a constrained optimization problem was formulated to determine the optimal protocols that enhance relevant metrics within clinically relevant domains. The practicality of the results was demonstrated through evaluation in real clinical settings and comparing the optimal protocols with the suggested protocol via routine clinical practice guidelines.

The accurate ground truth images and associated renditions were employed to train DNNs for harmonization aims. Different DNNs were tailored to the specific domain specifications, modalities, and available information. To enhance the generalizability of the developed DNNs, supplementary information from the physics of the acquisition or regulatory loss stemming from the harmonization principle was introduced and implemented. The algorithms were tested on diverse clinical data for various tasks, enabling the evaluation of quantification and the measurement of feature deviations.