Interactive medical image segmentation - towards integrating human guidance and deep learning
<p>Medical image segmentation is an essential step in many clinical workflows involving diagnostics and patient treatment planning. Deep learning has advanced the field of medical image segmentation, particularly with respect to automating contouring. However, some anatomical structures, such...
| Κύριος συγγραφέας: | |
|---|---|
| Άλλοι συγγραφείς: | |
| Μορφή: | Thesis |
| Γλώσσα: | English |
| Έκδοση: |
2023
|
| _version_ | 1826312014858813440 |
|---|---|
| author | Trimpl, MJ |
| author2 | Gooding, MJ |
| author_facet | Gooding, MJ Trimpl, MJ |
| author_sort | Trimpl, MJ |
| collection | OXFORD |
| description | <p>Medical image segmentation is an essential step in many clinical workflows involving diagnostics and patient treatment planning. Deep learning has advanced the field of medical image segmentation, particularly with respect to automating contouring. However, some anatomical structures, such as tumours, are challenging for fully automated methods. When automatic methods fail, manual contouring is required. In such cases, semi-automatic tools can support clinicians in contouring tasks. The objective of this thesis was to leverage clinicians’ expert knowledge when performing segmentation tasks, allowing for interactions along the segmentation workflow and improving deep learning predictions. </p>
<p>In this thesis, a deep learning approach is proposed that produces a 3D segmentation of a structure of interest based on a user-provided input. If trained on a diverse set of structures, state-of-the-art performance was achieved for structures included in the training set. More importantly, the model was also able to generalize and make predictions for unseen structures that were not represented in the training set. Various avenues to guide user interaction and leverage multiple user inputs more effectively were also investigated. These further improved the segmentation performance and demonstrated the ability to accurately segment a broad range of anatomical structures. </p>
<p>An evaluation by clinicians demonstrated that time spent contouring was reduced when using the contextual deep learning tool as compared to conventional contouring tools. This evaluation also revealed that the majority of contouring time is observation time, which is only indirectly affected by the segmentation approach. This suggests, that user interface design and guiding the user’s attention to critical areas can have a large impact on time taken on the contouring task. </p>
<p>Overall, this thesis proposes an interactive deep learning segmentation method, demonstrates its clinical impact, and highlights the potential synergies between clinicians and artificial intelligence.</p> |
| first_indexed | 2024-03-07T08:21:21Z |
| format | Thesis |
| id | oxford-uuid:3905a490-abef-44a7-8c6a-f392822b3b91 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T08:21:21Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:3905a490-abef-44a7-8c6a-f392822b3b912024-02-05T10:42:22ZInteractive medical image segmentation - towards integrating human guidance and deep learningThesishttp://purl.org/coar/resource_type/c_db06uuid:3905a490-abef-44a7-8c6a-f392822b3b91EnglishHyrax Deposit2023Trimpl, MJGooding, MJStride, EVallis, K<p>Medical image segmentation is an essential step in many clinical workflows involving diagnostics and patient treatment planning. Deep learning has advanced the field of medical image segmentation, particularly with respect to automating contouring. However, some anatomical structures, such as tumours, are challenging for fully automated methods. When automatic methods fail, manual contouring is required. In such cases, semi-automatic tools can support clinicians in contouring tasks. The objective of this thesis was to leverage clinicians’ expert knowledge when performing segmentation tasks, allowing for interactions along the segmentation workflow and improving deep learning predictions. </p> <p>In this thesis, a deep learning approach is proposed that produces a 3D segmentation of a structure of interest based on a user-provided input. If trained on a diverse set of structures, state-of-the-art performance was achieved for structures included in the training set. More importantly, the model was also able to generalize and make predictions for unseen structures that were not represented in the training set. Various avenues to guide user interaction and leverage multiple user inputs more effectively were also investigated. These further improved the segmentation performance and demonstrated the ability to accurately segment a broad range of anatomical structures. </p> <p>An evaluation by clinicians demonstrated that time spent contouring was reduced when using the contextual deep learning tool as compared to conventional contouring tools. This evaluation also revealed that the majority of contouring time is observation time, which is only indirectly affected by the segmentation approach. This suggests, that user interface design and guiding the user’s attention to critical areas can have a large impact on time taken on the contouring task. </p> <p>Overall, this thesis proposes an interactive deep learning segmentation method, demonstrates its clinical impact, and highlights the potential synergies between clinicians and artificial intelligence.</p> |
| spellingShingle | Trimpl, MJ Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title | Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title_full | Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title_fullStr | Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title_full_unstemmed | Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title_short | Interactive medical image segmentation - towards integrating human guidance and deep learning |
| title_sort | interactive medical image segmentation towards integrating human guidance and deep learning |
| work_keys_str_mv | AT trimplmj interactivemedicalimagesegmentationtowardsintegratinghumanguidanceanddeeplearning |