Predicting the future location of lymph nodes in rectal cancer patient during treatment using scans acquired prior to treatment
Author
Andresen, Anne Haahr
Term
4. term
Publication year
2022
Pages
79
Abstract
Rectal cancer is common, and recent treatment strategies sometimes omit surgery in favor of combined chemo-radiotherapy. To increase effectiveness while sparing healthy tissue, radiation fields should be reduced, which requires understanding motion of targets such as lymph nodes during treatment. Because treatment-phase imaging is typically unavailable, this thesis investigates whether future lymph node locations can be predicted from pre-treatment scans and derived predictors. The project developed a pipeline comprising (1) identification and quantification of literature-informed predictors, (2) data augmentation using a stacked autoencoder to expand the dataset, and (3) a deep learning regression model with a multi-branch network architecture to estimate subsequent lymph node positions. Validation suggested the selected predictors were plausible and training was stable. However, the final model yielded relatively high mean squared error across all outputs, indicating that further development of the model and data pipeline is needed to achieve clinically useful predictive accuracy.
Rektumkræft er en hyppig sygdom, hvor nyere behandlingsstrategier i nogle tilfælde udelader operation og i stedet kombinerer kemo- og stråleterapi. For at øge effekten uden at skade raskt væv ønskes strålefeltet indsnævret, men det kræver viden om, hvordan målområdet – herunder lymfeknuder – bevæger sig under behandlingen. Da der normalt ikke optages skanninger under forløbet, undersøger dette projekt, om den fremtidige placering af lymfeknuder kan forudsiges ud fra præbehandlingsskanninger og afledte prædiktorer. Projektet udviklede en pipeline med (1) identifikation og kvantificering af prædiktorer fra litteraturen, (2) dataaugmentation via en stacked autoencoder for at udvide datasættet og (3) en dyb lærings-regressionsmodel designet som en flergrebet netværksarkitektur, der estimerer lymfeknuders efterfølgende position. Valideringen indikerede, at de udvalgte prædiktorer var meningsfulde, og træningen var stabil. Den endelige model opnåede dog en relativt høj mean squared error for alle udgange, hvilket peger på, at yderligere model- og datapipeline-udvikling er nødvendig for at opnå klinisk relevant prædiktionsnøjagtighed.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords