Naji, Hussein (2024). Deep Learning for Prediction of Recurrence of Diffuse Large B-Cell Lymphoma. PhD thesis, Universität zu Köln.
PDF
Dissertation_Naji.pdf Download (68MB) |
Abstract
Diffuse large B-cell lymphoma (DLBCL) is an aggressive and highly heterogeneous tumor which is fatal without treatment. It is the most common non-Hodgkin lymphoma in adults. Up to today, its treatment presents a serious challenge as up to half of the treated patients suffer from recurrence of this cancer. Therefore, an early detection of those patients who do not benefit from standard first-line treatments would be of high significance as treatment procedures could be more personalized to prevent a recurrence. Previous research suggested that clinical data, gene expression markers, and genetic abnormalities are strong prognostic factors for recurrence of various cancer types. Recently, deep learning (DL) has shown increasing success in predicting recurrence of cancer from histological images. DL models, despite being highly accurate predictors, are often criticized for their black box characteristic. However, state-of-the-art research, such as models that use multiple instance learning, attention-based mechanisms, and/or class activation mapping methods, indicates that it is possible to make models more explainable. From a biological perspective, the morphology of tumor cells has also been used for several medical prediction and classification tasks because it is crucial to grasp a full picture of the tumor microenvironment. All experiments conducted within this project are based on histological images of DLBCL. Firstly, it is shown that morphological features extracted from cell nuclei are strong indicators to distinguish between different cell types in DLBCL tumors. Secondly, cell nuclei are automatically segmented and classified to close the gap of non-existing segmentation approaches for lymphoma images. The last part deals with the prediction of recurrence of DLBCL and combines the methods and findings of the previous approaches with attention-based mechanisms of DL models to biologically interpret the model’s prediction decision. To this end, it was not only possible to distinguish nonrecurred from recurred patients but also to explain which morphological features of cell nuclei were indicative for the prediction task. This allows for both, an upfront detection of recurred patients and to find distinct morphological patterns between both groups of interest.
Item Type: | Thesis (PhD thesis) | ||||||||
Translated title: |
|
||||||||
Translated abstract: |
|
||||||||
Creators: |
|
||||||||
URN: | urn:nbn:de:hbz:38-732294 | ||||||||
Date: | 2024 | ||||||||
Language: | English | ||||||||
Faculty: | Faculty of Mathematics and Natural Sciences | ||||||||
Divisions: | Faculty of Mathematics and Natural Sciences > Department of Mathematics and Computer Science > Institute of Computer Science | ||||||||
Subjects: | Data processing Computer science Medical sciences Medicine |
||||||||
Uncontrolled Keywords: |
|
||||||||
Date of oral exam: | 12 July 2024 | ||||||||
Referee: |
|
||||||||
Refereed: | Yes | ||||||||
URI: | http://kups.ub.uni-koeln.de/id/eprint/73229 |
Downloads
Downloads per month over past year
Export
Actions (login required)
View Item |