AI for emphysema detection (storage)

NAISS 2023/6-336


NAISS Medium Storage

Principal Investigator:

Paolo Soda


UmeƄ universitet

Start Date:


End Date:


Primary Classification:

20603: Medical Image Processing

Secondary Classification:

10201: Computer Sciences




To date, we are witnessing a widespread and growing adoption of Artificial Intelligence (AI) in several fields, healthcare included. Advancements in deep learning (DL) should make significant contributions to boost diagnosis, prognosis, and treatment decisions. This project delivers an AI-based solution for pulmonary emphysema detection. Early diagnosis will be facilitated as well as better intervention guidance for the clinicians. Emphysema is a hallmark of Chronic Obstructive Pulmonary Disease, a leading cause of early death worldwide. Moreover, it is also an independent risk factor for lung cancer. In recent decades, Computed Tomography (CT) has been used as the main diagnostic platform for the detection and quantification of emphysema. In clinical practice, the quantitative assessment identifies emphysema as low attenuation areas under a specific cut-off threshold, commonly set to -950 HU. Despite its wide adoption, this method lacks consensus on an optimal cut-off threshold and is prone to measurement variation, asking for new solutions that can encompass this limitation. Moreover, CT is associated with a risk of ionizing radiations making the use of low-dose CT (LDCT) highly recommended by medical societies. However, in LDCT the overall image quality decreases, compromising disease assessment and increasing uncertainty in the diagnosis. We hypothesize that AI can help overcome these limitations: hence, we aim to deepen the potential of AI for emphysema identification by focusing on three main aspects: enhancing the image quality of LDCT images, developing an explainable decision support system for emphysema quantification, and integrating both aspects into an end-to-end trainable framework. In detail, our three goals are: first, to enhance image quality, a task usually referred to as image denoising, to reduce uncertainties in the images, making the quality of low-dose CT comparable with standard CT scans. By leveraging data-driven approaches, such as those based on generative deep neural networks, denoising will be faced as an Image-to-Image translation task. Second, to develop an explainable AI-based decision support system, which can detect emphysema and also detect the area of the lungs where the emphysema is located but also provide explanations about how much an input feature contributes to the final predictions. Third, although both AI-based denoising and computer-aided detection/classification have been extensively studied, they are often treated separately and sequentially according to their respective goals, neglecting the interplay between noise reduction and detection/classification. For this reason, our largest efforts will be directed toward the investigation of unified approaches for jointly optimizing the denoising and classification problem in an end-to-end fashion. To this end, a well-designed deep learning framework will be developed by leveraging techniques such as latent space manipulation and custom loss function optimization. We will develop an extensive validation using data available within the SCAPIS pilot study. Moreover, we plan to include publicly available datasets, such as the Mayo Clinic dataset which contains chest CT scans acquired at routine dose level and simulated reduced dose.