Publicação
Fibrosis segmentation in cardiac resonance imaging
| Resumo: | Heart disease, a leading cause of mortality worldwide, is associated in many cases with a condition known as myocardial fibrosis, assuming a crucial prognostic in disease assessment. Fibrosis manifests in two forms, interstitial and replacement, both affect the heart tissue, however, the first may be reversible with early identification and intervention, and the second causes permanent scar tissue. Given this, the patient needs a precise diagnostic for effective management and treatment, so accurate fibrous tissue localization is needed, to reduce complications according to the patient’s condition. Manual segmentation is the gold standard for accuracy, although it’s time-consuming and costly. For this reason, automated methods have been developed to improve speed and repeatability compared to manual methods, however, the accuracy is still not reliable, due to difficulties in segmenting small and irregular tissues, resolution challenges, and cardiac magnetic resonance imaging quality . Considering the main problem, this project aims to introduce a new attention mechanism to improve an algorithm that can automatically segment the left ventricular myocardium and distinguish between healthy and fibrous tissue using advanced deep-learning methods. Regarding the methodology, we will have the method divided into two phases. In the first stage, a 2D U-Net with Balanced Steady-State Free Precession (bSSFP), Late Gadolinium Enhancement (LGE), and T2 sequences was used as input, to segment three key anatomical structures of the heart: left ventricle, right ventricle, and the left ventricular myocardium, to predict the myocardium area and identify possible locations for edema and scar. In the second stage, the left ventricular myocardium mask from the first stage, accompanied by the same three magnetic resonance imaging sequences and two new images, the LGE-myo and T2-myo, will be used to refine the segmentation and detection of the lesion. In this stage, a new attention mechanism, Bilateral Local Attention (BLA), was implemented within the transformer encoder from the network variant U-Shape Nested Transformer (UNesT). This network comprises a Transformer-based encoder (Nested Hierarchical Transformer), and a convolution-based decoder. The dataset employed in this approach was provided by the MICCAI2020 MyoPS challenge and consists of 45 sets of cardiac magnetic resonance images with three sequences (bSSFP, LGE, T2). Contrary to expected, the results achieved did not significantly exceed those of state-of-the-art methods, although they remain competitive compared to the models that did not employ an ensemble approach, it was positioned in third place among the six methods that did not use this strategy. The average Dice Similarity Coefficient (DSC) of the method was 0.678, in terms of scar segmentation the DSC resulted was 0.640 ± 0.232, and for the combined scar and edema the DSC was 0.715 ± 0.110. In conclusion, the results suggest that the implemented attention mechanism, although promising, did not bring significant benefits in capturing detailed information from cardiac magnetic resonance images compared to other works. This may be attributed to challenges in adapting the model to the specific anatomical variability present in the analyzed images or the need for further investigation to fine-tune the hyperparameters of the second stage. |
|---|---|
| Autores principais: | Ferreira, Luis Filipe da Silva |
| Assunto: | Myocardial fibrosis Cardiac magnetic resonance imaging Deep learning Transformers Fibrose miocárdica Imagem de ressonância magnética cardíaca Engenharia e Tecnologia::Outras Engenharias e Tecnologias |
| Ano: | 2024 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso restrito |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Heart disease, a leading cause of mortality worldwide, is associated in many cases with a condition known as myocardial fibrosis, assuming a crucial prognostic in disease assessment. Fibrosis manifests in two forms, interstitial and replacement, both affect the heart tissue, however, the first may be reversible with early identification and intervention, and the second causes permanent scar tissue. Given this, the patient needs a precise diagnostic for effective management and treatment, so accurate fibrous tissue localization is needed, to reduce complications according to the patient’s condition. Manual segmentation is the gold standard for accuracy, although it’s time-consuming and costly. For this reason, automated methods have been developed to improve speed and repeatability compared to manual methods, however, the accuracy is still not reliable, due to difficulties in segmenting small and irregular tissues, resolution challenges, and cardiac magnetic resonance imaging quality . Considering the main problem, this project aims to introduce a new attention mechanism to improve an algorithm that can automatically segment the left ventricular myocardium and distinguish between healthy and fibrous tissue using advanced deep-learning methods. Regarding the methodology, we will have the method divided into two phases. In the first stage, a 2D U-Net with Balanced Steady-State Free Precession (bSSFP), Late Gadolinium Enhancement (LGE), and T2 sequences was used as input, to segment three key anatomical structures of the heart: left ventricle, right ventricle, and the left ventricular myocardium, to predict the myocardium area and identify possible locations for edema and scar. In the second stage, the left ventricular myocardium mask from the first stage, accompanied by the same three magnetic resonance imaging sequences and two new images, the LGE-myo and T2-myo, will be used to refine the segmentation and detection of the lesion. In this stage, a new attention mechanism, Bilateral Local Attention (BLA), was implemented within the transformer encoder from the network variant U-Shape Nested Transformer (UNesT). This network comprises a Transformer-based encoder (Nested Hierarchical Transformer), and a convolution-based decoder. The dataset employed in this approach was provided by the MICCAI2020 MyoPS challenge and consists of 45 sets of cardiac magnetic resonance images with three sequences (bSSFP, LGE, T2). Contrary to expected, the results achieved did not significantly exceed those of state-of-the-art methods, although they remain competitive compared to the models that did not employ an ensemble approach, it was positioned in third place among the six methods that did not use this strategy. The average Dice Similarity Coefficient (DSC) of the method was 0.678, in terms of scar segmentation the DSC resulted was 0.640 ± 0.232, and for the combined scar and edema the DSC was 0.715 ± 0.110. In conclusion, the results suggest that the implemented attention mechanism, although promising, did not bring significant benefits in capturing detailed information from cardiac magnetic resonance images compared to other works. This may be attributed to challenges in adapting the model to the specific anatomical variability present in the analyzed images or the need for further investigation to fine-tune the hyperparameters of the second stage. |
|---|