IMAGE BASED FEATURE EXTRACTION TO IMPROVE SURVIVAL ANALYSIS IN HEAD AND NECK CANCER

Abstract

In this thesis we performed a pooled cohort study to investigate the role
of radiomics in head and neck cancer prognosis. The aim was to investigate prognostic value for overall survival and cancer recurrence of radiomics
combined with previously studied demographic and clinical risk factors. Radiomics features were extracted from the gross tumor volume on a computed
tomography captured prior to radiotherapy treatment. Both standard statistical model such as Cox regression, and common machine learning methods
such as random survival forest, DeepSurv and DeepHit, were used. The cancer type was constricted to oropharyngeal head and neck cancer due to large
amount of missing data in the other head and neck cancer types. Prognostic
performance for local recurrence was improved using shape related radiomics
(sphericity) and clustering based methods (PCA). In contrast, the results
showed no improved performance for overall survival (OS) for any model,
where a possible reason might be too few events per covariate or that OS
depends mainly on factors not captured by the radiomics data.
These results indicate a role for radiomics in prognostic evaluation, which
could prove to be useful treatment decision making and research guidance.

Popular Abstract

More people get cancer today than ever before, and according to the World
Health Organization (WHO) the trend in cancer cases will keep rising in the
near future with an increase of 77 % from 2020 to 2050. Despite the increase
in cancer cases, cancer mortality seems to be decreasing. One reason that is
commonly given for the decrease in mortality is improved treatment methods.
Decisions on treatment are made through patient specific information (such
as age or overall health status) and tumor specific information (such as cancer
type). Tumor specific information can be obtained through biopsy or medical
imaging. One way to quantify information in medical images is radiomics,
which is a method that is currently not used to inform treatment decision
making.
Radiomics extracts information from images with the aim to describe the
region of interest in the image (such as the tumor) with great detail. The
aim of this thesis was to use survival analysis to evaluate if radiomics could
help in predicting survival and recurrence in head and neck cancer, with the
goal to potentially use radiomics in treatment decision making in the future.
To do this, both standard statistical methods such as Cox regression, and
machine learning methods such as DeepSurv, DeepHit and Random survival
forest, were used. These models were used to model the probability of getting
the event (either death or cancer recurrence) over time. The machine learning
methods were only used to model death probability.
The results showed that neither standard statistical models or machine
learning models could utilize the radiomics to improve the probability predictions for the case of death. Radiomics did seem to provide an improvement
for recurrence prediction. There therefore seems to be promise to use the
radiomic information to inform on how treatment decisions are made.

• author: Linnér, Anton
• supervisor: Anna Lindgren; André Haraldsson
• organization: Mathematical Statistics
• course: MASM02 20241
• year: 2024
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Keywords: Radiomics, Head and neck cancer, Oropharynx, Cox regression, DeepSurv, DeepHit, Random survival forest, Overall survival, Local recurrence
• publication/series: Master’s Theses in Mathematical Sciences
• report number: LUNFMS-3130-2024
• ISSN: 1404-6342
• other publication id: 2024:E70
• language: English
• id: 9175924
• date added to LUP: 2024-10-03 15:25:40
• date last changed: 2024-10-03 15:25:40