Accurate and efficient segmentation of malignant lymphomas on fluorodeoxyglucose
positron emission tomography/computed tomography (FDG PET/CT) images
would help to speed up and automate the segmentation process. This thesis explores
deep learning method nnUNet in the context of lymphoma image segmentation.
FDG PET/CT images of 202 patients were used in this thesis. The images were
used to train 9 different models with different training sets and the duration’s of
learning. We were able to satisfactorily segment areas of lymphomas with trained
models, where larger areas of lymphoma involvement were better segmented by the
models while smaller lesions were associated with worse segmentation results. The
model that stood out was the one with the largest training set of 143 patients and
the duration of training of 10 h, which was the second longest training time, being
also comparable in terms of Dice metric values to other tumour segmentation models
found in the literature.
The findings of this research highlight potential of AI in clinical work and illustrate
the promising features of deep learning methods such as nnUNet. As the
field of medical image analysis continues to evolve, the integration of advanced AI
techniques promises more personalised and efficient healthcare solutions.
|
