Jury

Summary:

Cancer is one of the leading causes of death worldwide, with almost 10 million deaths in 2022, including 1.5 million in Europe. Non-Hodgkin's lymphomas (NHL) are cancers of the immune system and are responsible for more than 250,000 deaths in 2022. In France, the incidence of NHL rose by 21% between 1998 and 2017, and projections predict a significant increase in cases and deaths between now and 2045. Diffuse large B-cell lymphoma (DLBCL) is the most common type of NHL, with two molecular subtypes identified using gene expression profiling techniques: ABC and GCB, the former of which is characterised by a lower survival rate and the need for specific treatment. Several methods exist for determining the LDGC-B subtype, including RT-MLPA (molecular technique), immunohistochemistry and examination of haematoxylin-eosin-stained tissue. However, these methods have limitations in terms of cost, time and accuracy. In this thesis, we propose to investigate the potential of machine learning and deep learning-based methods to improve the diagnosis of LDGC-B patients, both in terms of molecular subtyping and in predicting response to treatment. Using the acquisition of high-resolution histopathological images, we propose two methodologies leading to models capable of predicting a patient's molecular subtype and response to treatment on the basis of these images. The experimental results of this work have shown the potential contribution that these methods can make to the diagnosis of LDGC-B, but also their ability to be generalised to other types of cancer and prediction tasks.

Abstract:

Cancer is one of the leading causes of death worldwide, with almost 10 million deaths in 2022, including 1.5 million in Europe. Non-Hodgkin's lymphomas (NHL) are cancers of the immune system, and are responsible for more than 250,000 deaths in 2022. In France, the incidence of NHL rose by 21% between 1998 and 2017, and projections predict a significant increase in cases and deaths by 2045. Diffuse large B-cell lymphoma (DLBCL) is the most common type of NHL, with two molecular subtypes identified using gene expression profiling techniques: ABC and GCB, the former of which is characterized by a lower survival rate and the need for specific treatments. Several methods exist to determine the DLBCL subtype, including RT-MLPA (molecular technique), immunohistochemistry and examination of hematoxylin-eosin-stained tissue. Nevertheless, these methods have limitations in terms of cost, time and accuracy. In this thesis, we propose to investigate the potential of machine learning and deep learning-based methods to improve the diagnosis of patients with DLBCL, both in terms of molecular subtyping but also for the prediction of treatment response. Thanks to the acquisition of high-resolution histopathological images, we propose two methodologies leading to models capable of predicting a patient's molecular subtype and treatment response from these images. The experimental results of these works have demonstrated the potential contribution of these methods to the diagnosis of DLBCL, but also their ability to be generalized to other types of cancer and predictive tasks.