Segmentação de estruturas pulmonares e lesões na tomografia computadorizada com dados heterogêneos

Author(s)

Diedre Santos do Carmo

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação
• Defense date: 2024-03-27

Examining board members

Roberto de Alencar Lotufo; Roberto Medeiros de Souza; Marco Antônio Gutierrez; Luis Fernando Gomez Gonzalez

• Advisor: Leticia Rittner; Roberto de Alencar Lotufo

Abstract

The response to the COVID-19 pandemic has highlighted the potential of deep learning methods in facilitating the diagnosis and prognosis of lung diseases through automated segmentation of normal and abnormal tissue on computed tomography (CT). Such methods not only have the potential to assist in clinical decision-making but also contribute to the understanding of new diseases. Given the intensive nature of manual segmentation for large chest CT sets, there is an urgent need for reliable automated approaches that enable efficient analysis of chest CT anatomy in vast research databases, especially in less frequently annotated structures such as parenchymal consolidations and ground-glass opacities caused by pneumonia. The development of methods that deal with these particular findings is hampered by the scarce availability of manually annotated data. To address this problem, we propose polymorphic multitask learning (PML). PML is used to optimize a network with a fixed number of output channels to represent multiple hierarchical anatomical structures, indirectly optimizing more complex structures with simpler annotations. Additionally, we use multitask learning with multiple segmentation heads to develop MEDPSeg, a method for segmenting lungs, airways, pulmonary artery, and lung lesions with separation of lesion into ground-glass opacities and parenchymal consolidations, all in a single prediction. More than 6000 CT scans containing varied manual and automatic annotation formats from different sources were used in the development of the proposed method. We show that the proposed APM training technique results in state-of-the-art multitasking performance for MEDPSeg, with Dice improvements of up to 15% in the segmentation of 5 different tasks when compared to target specialized versions. Furthermore, the parenchymal consolidation segmentation performance from MEDPSeg is superior to the current state-of-the-art. Finally, we provide an open source implementation with a graphical interface at https://github.com/MICLab-Unicamp/medpseg, capable of providing full volumetric results on high-quality CT scans in 1 minute with a NVIDIA 2060 8GB graphics processing uni (AU)

• FAPESP's process: 19/21964-4 - Lung diseases and COVID diagnosis from computed tomography images using explainable deep learning
• Grantee: Diedre Santos do Carmo
• Support Opportunities: Scholarships in Brazil - Doctorate