| _version_ | 1866902168422842368 |
|---|---|
| author | ALBEASH, Ali HABERT, Paul GAUBERT, Jean-Yves DUFOUR, Jean-Charles JACQUIER, Alexis |
| author_facet | ALBEASH, Ali HABERT, Paul GAUBERT, Jean-Yves DUFOUR, Jean-Charles JACQUIER, Alexis |
| contents | <p>This master’s thesis presents a comprehensive investigation into the application of deep learning methodologies for the fine-grained segmentation of the 20 bronchopulmonary segments of the lungs from computed tomography (CT) scans. The work specifically addresses a persistent challenge in medical image analysis: the accurate delineation of anatomically complex and small-scale pulmonary structures, which are often difficult to segment using conventional or coarse-grained approaches. By focusing on bronchopulmonary segmentation at a segmental level—rather than whole-lung or lobar segmentation—the study contributes to advancing the level of anatomical precision achievable through artificial intelligence.</p> <p>The research is based on a carefully curated dataset comprising 89 CT scans, each associated with high-quality three-dimensional segmentation masks. These annotations were initially produced through meticulous manual segmentation and subsequently validated by expert radiologists, ensuring the reliability of the ground truth data. The dataset includes both healthy and pathological cases, thereby incorporating anatomical variability and enhancing the robustness of the developed models. A structured data preparation pipeline was implemented, involving harmonization of imaging formats, verification of spatial alignment between images and masks, label standardization across all bronchopulmonary segments, and appropriate handling of volumetric data.</p> <p>Methodologically, the study adopts a comparative experimental design to evaluate the performance of different deep learning architectures. A baseline 2D U-Net model was first implemented to establish a reference point for segmentation performance. This was followed by the application of the nnU-Net framework, which introduces automated configuration and optimization strategies tailored to the dataset characteristics. Both 2D and 3D configurations of nnU-Net were explored, with particular emphasis on the 3D low-resolution variant, which leverages volumetric context while maintaining computational feasibility. The training procedures incorporated advanced techniques such as data normalization, one-hot encoding of segmentation masks, and five-fold cross-validation, ensuring a rigorous and reproducible evaluation process.</p> <p>The results demonstrate a clear performance gradient across the tested models, with the 3D nnU-Net architecture significantly outperforming its 2D counterparts. Quantitative evaluation using standard segmentation metrics, including the Dice similarity coefficient and Intersection over Union (IoU), revealed that the 3D model achieved mean scores of approximately 0.83 and 0.72, respectively. These improvements are primarily attributed to the model’s ability to capture three-dimensional spatial dependencies and contextual information across slices, which is particularly beneficial for accurately segmenting smaller and anatomically intricate lung segments. In contrast, the 2D models exhibited limitations in capturing inter-slice continuity, leading to reduced performance, especially in less represented or lower-volume segments.</p> <p>Beyond quantitative performance, the study underscores the clinical relevance of high-precision automated lung segmentation. Accurate delineation of bronchopulmonary segments has direct implications for a range of medical applications, including early disease detection, surgical planning (such as segmentectomy), radiotherapy targeting, and longitudinal monitoring of pulmonary conditions. By reducing reliance on time-consuming manual annotation and minimizing inter-observer variability, the proposed approach offers a pathway toward more efficient and standardized clinical workflows.</p> <p>Nevertheless, the study acknowledges several limitations. The relatively modest dataset size may constrain the generalizability of the findings, particularly across diverse populations and imaging protocols. Additionally, the segmentation of smaller bronchopulmonary segments remains inherently challenging due to class imbalance and limited representation in the training data. These factors highlight the need for future work focusing on dataset expansion, incorporation of more diverse clinical cases, and the development of specialized techniques—such as attention mechanisms or multi-scale modeling—to further enhance performance.</p> <p>In conclusion, this work provides a robust proof of concept for the application of adaptive deep learning frameworks, particularly nnU-Net, in complex multi-class medical image segmentation tasks. It demonstrates that leveraging three-dimensional contextual information is critical for achieving high levels of accuracy in detailed anatomical segmentation. The findings lay a strong foundation for future research aimed at improving model generalization, scalability, and clinical integration, ultimately contributing to the advancement of AI-driven solutions in medical imaging.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19685118 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Advanced Lung Bronchopulmonary Segmentation with Deep Learning ALBEASH, Ali HABERT, Paul GAUBERT, Jean-Yves DUFOUR, Jean-Charles JACQUIER, Alexis Image Processing, Computer-Assisted Artificial Intelligence Deep Learning Neural Networks, Computer Three-Dimensional Imaging <p>This master’s thesis presents a comprehensive investigation into the application of deep learning methodologies for the fine-grained segmentation of the 20 bronchopulmonary segments of the lungs from computed tomography (CT) scans. The work specifically addresses a persistent challenge in medical image analysis: the accurate delineation of anatomically complex and small-scale pulmonary structures, which are often difficult to segment using conventional or coarse-grained approaches. By focusing on bronchopulmonary segmentation at a segmental level—rather than whole-lung or lobar segmentation—the study contributes to advancing the level of anatomical precision achievable through artificial intelligence.</p> <p>The research is based on a carefully curated dataset comprising 89 CT scans, each associated with high-quality three-dimensional segmentation masks. These annotations were initially produced through meticulous manual segmentation and subsequently validated by expert radiologists, ensuring the reliability of the ground truth data. The dataset includes both healthy and pathological cases, thereby incorporating anatomical variability and enhancing the robustness of the developed models. A structured data preparation pipeline was implemented, involving harmonization of imaging formats, verification of spatial alignment between images and masks, label standardization across all bronchopulmonary segments, and appropriate handling of volumetric data.</p> <p>Methodologically, the study adopts a comparative experimental design to evaluate the performance of different deep learning architectures. A baseline 2D U-Net model was first implemented to establish a reference point for segmentation performance. This was followed by the application of the nnU-Net framework, which introduces automated configuration and optimization strategies tailored to the dataset characteristics. Both 2D and 3D configurations of nnU-Net were explored, with particular emphasis on the 3D low-resolution variant, which leverages volumetric context while maintaining computational feasibility. The training procedures incorporated advanced techniques such as data normalization, one-hot encoding of segmentation masks, and five-fold cross-validation, ensuring a rigorous and reproducible evaluation process.</p> <p>The results demonstrate a clear performance gradient across the tested models, with the 3D nnU-Net architecture significantly outperforming its 2D counterparts. Quantitative evaluation using standard segmentation metrics, including the Dice similarity coefficient and Intersection over Union (IoU), revealed that the 3D model achieved mean scores of approximately 0.83 and 0.72, respectively. These improvements are primarily attributed to the model’s ability to capture three-dimensional spatial dependencies and contextual information across slices, which is particularly beneficial for accurately segmenting smaller and anatomically intricate lung segments. In contrast, the 2D models exhibited limitations in capturing inter-slice continuity, leading to reduced performance, especially in less represented or lower-volume segments.</p> <p>Beyond quantitative performance, the study underscores the clinical relevance of high-precision automated lung segmentation. Accurate delineation of bronchopulmonary segments has direct implications for a range of medical applications, including early disease detection, surgical planning (such as segmentectomy), radiotherapy targeting, and longitudinal monitoring of pulmonary conditions. By reducing reliance on time-consuming manual annotation and minimizing inter-observer variability, the proposed approach offers a pathway toward more efficient and standardized clinical workflows.</p> <p>Nevertheless, the study acknowledges several limitations. The relatively modest dataset size may constrain the generalizability of the findings, particularly across diverse populations and imaging protocols. Additionally, the segmentation of smaller bronchopulmonary segments remains inherently challenging due to class imbalance and limited representation in the training data. These factors highlight the need for future work focusing on dataset expansion, incorporation of more diverse clinical cases, and the development of specialized techniques—such as attention mechanisms or multi-scale modeling—to further enhance performance.</p> <p>In conclusion, this work provides a robust proof of concept for the application of adaptive deep learning frameworks, particularly nnU-Net, in complex multi-class medical image segmentation tasks. It demonstrates that leveraging three-dimensional contextual information is critical for achieving high levels of accuracy in detailed anatomical segmentation. The findings lay a strong foundation for future research aimed at improving model generalization, scalability, and clinical integration, ultimately contributing to the advancement of AI-driven solutions in medical imaging.</p> |
| title | Advanced Lung Bronchopulmonary Segmentation with Deep Learning |
| topic | Image Processing, Computer-Assisted Artificial Intelligence Deep Learning Neural Networks, Computer Three-Dimensional Imaging |
| url | https://doi.org/10.5281/zenodo.19685118 |