University of Twente Student Theses
Automatic 3D segmentation of prolapsed bladders from low-field MRI using 3D U-Net
Wever, A.M. de (2023) Automatic 3D segmentation of prolapsed bladders from low-field MRI using 3D U-Net.
|
PDF
5MB |
| Abstract: | Anterior vaginal wall prolapse affects about 41% to 50% of women. Anterior Colporrhaphy is the most performed pelvic organ prolapse surgery. It has a 41% chance of operative failure and a 10% to 30% re-operation rate. There is currently no standardized technique among surgeons for colporrhaphy. A 3D model of the bladder could be useful for diagnosis, surgical planning and surgical evaluation. A deep learning model using a convolutional neural network in the form of a U-Net is trained to segment 3D models of all bladders (prolapsed and non-prolapsed) from MRI images. The data used is gathered mainly from the TORBO study at the University of Twente. After a selection of the data, aided by classification of the bladders in degree of prolapse, contrast and artefacts, there are 4 test scans and 29 training scans (24 training and 5 validation). A selection is also made for the dataset sizes of 15, 20, 25. All sizes are used for training to evaluate the effect of dataset size on the test mean DSC and HD. The best model is used to obtain predictions on the test scans. The network prediction is then used in Matlab to evaluate the bladders pre- and post-op in the PICS coordinate system to evaluate the effect of the surgery. Increasing the dataset size leads to an increase in test mean DSC and a decrease in the test mean HD. The model chosen to further evaluate the test scans has a DSC of 0.79 and a HD of 21.4. This model shows that the prolapsed bladders can be almost successfully segmented. Small extrusions in the prediction are still present and bladders strongly affected by bowel movement are not segmented correctly. Also the bladder wall is sometimes included in the prediction. With an increase of training data the model is expected to reach a mean DSCof around 0.85-0.90 which will further improve the results. The evaluation of the surgery can be done by evaluating the lowest point of the bladder from the 3D PICS coordinate system. It is concluded that the current model is a good step towards automated prolapsed bladder segmentation. |
| Item Type: | Essay (Bachelor) |
| Faculty: | TNW: Science and Technology |
| Programme: | Biomedical Technology BSc (56226) |
| Link to this item: | https://purl.utwente.nl/essays/98695 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page