Publication
Automatic quantitative assessment of Pectus Excavatum severity using CT images and deep learning
| Summary: | Pectus excavatum (PE) is the most common congenital abnormality of the thoracic cage, characterized by the inward displacement of the sternum and the adjacent costal cartilages. To assess and quantify the patient’s severity, three standard indices are extracted from computed tomography (CT) images: the Haller index; the correction index; and the asymmetry index. Currently, there is no solution to automate the extraction of these indices. The focus of this thesis is to develop an automatic system that analyzes the CT of the patient and, as a result, provides the measurements and indices of PE with accuracy and reproducibility, saving healthcare professionals’ time and increasing measurements’ reproducibility, which may ultimately improve patients’ diagnosis and care. Through the use of deep learning, a framework was developed consisting of three stages. In the first two stages, two convolution neural networks (CNNs), based on the UNet++ architecture, were used to learn the CT images’ features in order to select the optimal CT slice to take the measurements and identify the relevant eight keypoints. The final stage relies on geometric rules to calculate the measurements and indices of PE from the identified keypoints. Both automatic and manual analyses were performed on 269 CTs of 92 patients. For comparative purposes, intra-observer, inter-observer, and intra-patient variability of the estimated indices were analyzed in a subset of patients. The developed system proved to be viable, showing a good agreement with the manual approach, with small bias and narrow limits of agreement (LOA), being the results comparable with the inter-observer agreement. In the intra-patient analysis, the proposed framework outperformed the expert, showing a higher reproducibility between indices extracted from distinct CTs of the same patient. The developed system achieved an average relative error against manually extracted values of 4.09%, 5.59% and 1.91% for the Haller, correction and asymmetry indices, respectively. The results are promising, supporting the feasibility of using the developed framework for the automatic and accurate quantification of PE severity in a clinical context. |
|---|---|
| Main Authors: | Silva, Bruno André Pires da |
| Subject: | Pectus excavatum Computer-aided diagnosis Medical image analysis Deep learning Convolutional neural network Diagnóstico auxiliado por computador Análise de imagens médicas Aprendizagem profunda Rede neural convolucional Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática |
| Year: | 2021 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade do Minho |
| Language: | English |
| Origin: | RepositóriUM - Universidade do Minho |
| Summary: | Pectus excavatum (PE) is the most common congenital abnormality of the thoracic cage, characterized by the inward displacement of the sternum and the adjacent costal cartilages. To assess and quantify the patient’s severity, three standard indices are extracted from computed tomography (CT) images: the Haller index; the correction index; and the asymmetry index. Currently, there is no solution to automate the extraction of these indices. The focus of this thesis is to develop an automatic system that analyzes the CT of the patient and, as a result, provides the measurements and indices of PE with accuracy and reproducibility, saving healthcare professionals’ time and increasing measurements’ reproducibility, which may ultimately improve patients’ diagnosis and care. Through the use of deep learning, a framework was developed consisting of three stages. In the first two stages, two convolution neural networks (CNNs), based on the UNet++ architecture, were used to learn the CT images’ features in order to select the optimal CT slice to take the measurements and identify the relevant eight keypoints. The final stage relies on geometric rules to calculate the measurements and indices of PE from the identified keypoints. Both automatic and manual analyses were performed on 269 CTs of 92 patients. For comparative purposes, intra-observer, inter-observer, and intra-patient variability of the estimated indices were analyzed in a subset of patients. The developed system proved to be viable, showing a good agreement with the manual approach, with small bias and narrow limits of agreement (LOA), being the results comparable with the inter-observer agreement. In the intra-patient analysis, the proposed framework outperformed the expert, showing a higher reproducibility between indices extracted from distinct CTs of the same patient. The developed system achieved an average relative error against manually extracted values of 4.09%, 5.59% and 1.91% for the Haller, correction and asymmetry indices, respectively. The results are promising, supporting the feasibility of using the developed framework for the automatic and accurate quantification of PE severity in a clinical context. |
|---|
Funded activities
No funded projects found.