Artificial Count Enhancement of Lung Scintigraphic Images Using Deep Learning Techniques

Abstract

This thesis examines the use of artificial intelligence (AI) for enhancing photon counts in ventilation/perfusion (V/Q) scans. V/Q scintigraphy is essential in the nuclear medicine field for diagnosing pulmonary embolism (PE), a severe and potentially life-threatening condition. Traditional V/Q scanning methods are often lengthy, which not only cause discomfort for patients, and thus possibly affect the quality of the images obtained, but also has contributed to their declining use and the uprise of computed tomography pulmonary angiography (CTPA). This thesis aims to address these issues by proposing AI-based techniques that enhance low-count images in V/Q scans, with the aim of reducing scan times while maintaining accurate diagnoses.

The research began with a systematic review to assess the role of AI in V/Q scans. This review identified several promising areas for further research, including image enhancement, artifact removal, and the creation of pseudo-planar images from single-photon emission computed tomography (SPECT). This latter aspect is relevant since there has been research to support the transition from traditional planar imaging towards SPECT; thus, alternatives are needed that help physicians adapt.

A critical part of the research was evaluating the impact of various image resizing techniques on the noise characteristics inherent in V/Q images. This involved a comparative analysis of standard upsampling and downsampling methods, such as linear interpolation, against techniques designed to preserve Poisson counting statistics. These included linear interpolation followed by Poisson resampling for upsampling and sliding window summation for downsampling. Image quality was assessed using the structured similarity index (SSIM) and the logarithm of mean squared error (MSE), to compare resized images with those at the target resolution. The findings revealed that upsampling with Poisson resampling after linear interpolation yielded images consistent with correct photon count properties at the target resolution. In downsampling, while linear interpolation and sliding window summation were both effective at a 2x reduction, the latter method produced realistic images at a 4x reduction.

Furthermore, the thesis showcases the development, training, and validation of a deep learning model tailored for count enhancement using scintigraphic images. This model successfully created diagnostic quality images from simulated low-count images, which were derived from diagnostic-quality images at 10% counts, with full-count images serving as the ground truth. The effectiveness of various loss functions was determined through both qualitative and quantitative analyses. It was observed that a combination of L₁, perceptual, and adversarial losses yielded images most comparable to the ground truth. These AI-enhanced images underwent clinical evaluation by both experienced and novice V/Q scan readers. The assessment was based on three categories: low, moderate, and high similarity. The general consensus among the readers indicated a moderate to high similarity, suggesting that the AI-enhanced images retained essential diagnostic features.