Predicting Uterine Deformation Due to Applicator Insertion in Pre-Brachytherapy MRI Using Deep Learning

• Author / Creator

  Ghosh, Shrimanti

• In locally advanced cervical cancer (LACC), brachytherapy (BT) remains the gold standard for boosting to curative doses in radiotherapy. Progress towards balancing target and routine tissue dosimetry for better clinical outcomes has been made possible by magnetic resonance imaging (MRI)-guided imaging data for treatment planning and by improving traditional BT applicator geometry. However, further evolving the original “one size fits all” approach towards truly personalized BT delivery requires improvements along the entire BT care pathway. From the pre-insertion applicator selection to the post-insertion treatment plan optimization, many factors need to be optimized for each patient for the best possible tumour and morbidity outcomes. The expanding collection and sharing of data, increased computational power in machine learning (ML) and artificial intelligence (AI), deep learning (DL) are rapidly transforming society, and offer the potential for similar transformation within health care. The success of these algorithms is founded on their judicious capability for detecting complex patterns even in heterogeneous datasets.

  This thesis aims to develop ML and DL-based models and free-form deformation methods for building predictive models to guide BT processes and decisions for consistently better personalization in LACC. First, an automated segmentation algorithm is proposed to delineate the uterus from the background on MRI using a deep convolutional neural network (CNN) architecture (Inception-V4) along with auto-encoders. After automated uterus segmentation, a modified version of another deep CNN model i.e., U-net is utilized to predict the at-BT uterus shape from pre-BT MRI. Finally, a shape-based non-rigid registration (free form deformation) algorithm is proposed to measure or quantify the amount of complex and large deformations of the uterus structure due to BT applicator insertion. The study deals with the very challenging and complex problem of predicting the large anatomical deformations from pelvic MR-images due to BT applicator insertion. The proposed method achieved an average Dice Coefficient of 94.8\% and a Hausdorff distance of 3.06 mm, whereas the U-net yielded 92.4\% and 6.7 mm for the Dice score and Hausdorff Distance metrics, respectively in the uterus segmentation task. The quantitative evaluations demonstrated that the proposed method performed significantly better than U-net in terms of both Dice Coefficient and Hausdorff Distance. After that, a pre-trained modified U-net is proposed to predict the at-BT uterus position from only the pre-BT MRI. This method yielded an average Dice score of 89.5\% and a Hausdorff distance of 3.6 mm in predicting the uterine deformation automatically. Large anatomy deformations before and at the time of BT insertion were observed for most patients due to the insertion of the BT applicator. In order to quantify this deformation, a free-form deformation model-based non-rigid registration method is proposed. The applicator’s presence introduces a median uterine surface point-to-point displacement of 25.0 [10.0 - 62.5] mm and a median uterine cavity point-to-point displacement of 40.0 [12.0 - 68.0] mm from the pre-BT position.

  The challenge in implementing this algorithm was the inter-patient anatomical dissimilarity and extreme intra-patient uterine deformation from pre-BT to at-BT in the dataset. Increasing the size of our training dataset, with the inclusion of more heterogeneous images with anatomical variability, will improve the prediction accuracy of this DL-based algorithm. Our proposed DL-based model, despite being trained on heterogeneous and complex deformations, can successfully predict uterine distortion automatically due to applicator insertion using only the pre-BT MRI, which can guide the clinicians in selecting the most suitable applicator component and configuration ahead of the actual insertion procedure. These promise better, faster, and more streamlined clinical/technical decision-making before BT applicator insertion and plan optimization, potentially enabling more consistent application of BT personalization for LACC and improved dosimetric outcomes.

• Subjects / Keywords

  • Uterine deformation
  • cervical cancer
  • brachytherapy
  • deep learning
  • artificial intelligence
  • medical imaging

• Graduation date

  Spring 2023

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-zyqb-h385

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.