Abstract
This study applies computer vision and machine learning techniques to segment subglot-tic stenosis, a condition characterized by narrowing the airways below the vocal cords. The size of the airway decreases, and the remaining opening is called the lumen. The area be-tween the airways and the lumen is called stenosis. The goal is to develop an automated and accurate method for quantifying the extent of stenosis based on medical imaging, which is crucial for effective diagnosis and treatment planning.This study uses a set of medical images of patients’ subglottic stenosis. Despite a large number of images, only a small number are suitable for inclusion in a training dataset, and because there are only images, a suitable dataset must be built from these images. Various computer vision techniques are explored to determine the most suitable for building an an-notated dataset. The most suitable one, in this case, is binary thresholding, which is used for lumen segmentation. Stenosis segmentation is more difficult. The shape of the first tracheal ring is used to determine the stenosis. Validating these segmented parts caused problems.
Three machine learning models (Unet, DeepLabV3+, TransUnet) are selected to be trained on the dataset and compared to see which performs best. A grid search is used to find an optimal combination of batch sizes and learning rates at which the model performs best.
Performance is assessed using metrics such as Intersection over Union (IoU) and accu-racy. The model’s segmentation output is compared with the annotated ground truth, and the computational burden is also included in the assessment.
The results show a substantial difference in performance between segmenting the lu-men and segmenting the stenosis. All models show better results when segmenting lu-men than when segmenting stenosis. Lumen segmentation appears to be very good, while stenosis segmentation lags behind. TransUnet performs approximately the same as the other models but at a much higher computational cost.
The study shows that machine learning can effectively segment the lumen of subglottic stenosis from medical images. Stenosis segmentation requires more attention, especially when validating the dataset. The developed models do not yet offer doctors a reliable and efficient instrument. Future work should focus on expanding the dataset and validating it to improve the quality of the stenosis segmentation.
Date of Award | 20 Aug 2024 |
---|---|
Original language | English |
Supervisor | Gideon Maillette de Buij Wenniger (Examiner) & Arjen Hommersom (Co-assessor) |
Master's Degree
- Master Computer Science