Publicação
Deep Learning Approach for the Segmentation of Spinal Structures in Ultrasound Images
| Resumo: | Scoliosis, characterized by a three-dimensional deformation of the spine with a curvature exceeding 10° in the coronal plane, presents a diagnostic challenge. Conventionally, the gold standard for scoliosis diagnosis and spinal curvature measurement relies on X-ray imaging and the calculation of Cobb Angles. However, the ionizing radiation associated with X-rays poses potential risks, including the development of cancer. Moreover, the frequent use of X-rays for monitoring scoliosis progression or treatment outcomes increases the cumulative risk of these adverse effects. In contrast, ultrasound imaging emerges as a promising alternative. It is devoid of ionizing radiation, cost-effective, and highly portable, offering the potential for increased scoliosis screening and detection. Nonetheless, ultrasound spine imaging encounters challenges, notably low contrast and speckled noise, which compromise image quality. In this research endeavour, the application of cutting-edge deep learning techniques to address the challenges present in ultrasound images was studied. A novel deep learning model, known as the Hybrid Attention Recurrent Residual U-Net (Hybrid R2AU-Net), is proposed. It is specifically designed for segmenting spine structures in ultrasound spine images. The Hybrid R2AU-Net is a five-layer deep learning model that includes dense connections, Recurrent Residual Blocks (RecRes Blocks), and attention mechanisms in its architecture. The outcomes of this study were highly encouraging. The Hybrid R2AU-Net outperformed alternative deep learning models, delivering superior results with a Dice Score of 85.2%, a Jaccard Index of 74.3%, and a Detection Rate of 92.4%. In the future, these remarkable achievements underscore the potential of the Hybrid R2AU-Net to be seamlessly integrated into an automated scoliosis diagnosis system, promising a more radiation-free and efficient approach to scoliosis management. |
|---|---|
| Autores principais: | Monteiro, Maria Leonor Cohen |
| Assunto: | Scoliosis Ultrasound Segmentation Deep Learning Convolutional Neural Network |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Scoliosis, characterized by a three-dimensional deformation of the spine with a curvature exceeding 10° in the coronal plane, presents a diagnostic challenge. Conventionally, the gold standard for scoliosis diagnosis and spinal curvature measurement relies on X-ray imaging and the calculation of Cobb Angles. However, the ionizing radiation associated with X-rays poses potential risks, including the development of cancer. Moreover, the frequent use of X-rays for monitoring scoliosis progression or treatment outcomes increases the cumulative risk of these adverse effects. In contrast, ultrasound imaging emerges as a promising alternative. It is devoid of ionizing radiation, cost-effective, and highly portable, offering the potential for increased scoliosis screening and detection. Nonetheless, ultrasound spine imaging encounters challenges, notably low contrast and speckled noise, which compromise image quality. In this research endeavour, the application of cutting-edge deep learning techniques to address the challenges present in ultrasound images was studied. A novel deep learning model, known as the Hybrid Attention Recurrent Residual U-Net (Hybrid R2AU-Net), is proposed. It is specifically designed for segmenting spine structures in ultrasound spine images. The Hybrid R2AU-Net is a five-layer deep learning model that includes dense connections, Recurrent Residual Blocks (RecRes Blocks), and attention mechanisms in its architecture. The outcomes of this study were highly encouraging. The Hybrid R2AU-Net outperformed alternative deep learning models, delivering superior results with a Dice Score of 85.2%, a Jaccard Index of 74.3%, and a Detection Rate of 92.4%. In the future, these remarkable achievements underscore the potential of the Hybrid R2AU-Net to be seamlessly integrated into an automated scoliosis diagnosis system, promising a more radiation-free and efficient approach to scoliosis management. |
|---|