Self-Supervised Representation Learning for Survival Prediction from Multichannel Microscopy Images of Cancer Tissue

Sammanfattning

Lung cancer is a major global health challenge and one of the deadliest and most common cancer types in humans [57]. Although recent advances in molecular profiling, targeted therapies, immunotherapy, and early screening have improved the management of non-small cell lung cancer (NSCLC), lung cancer remains the leading cause of cancer-related deaths worldwide [26]. In 2022, lung cancer was reported as the most frequently diagnosed cancer worldwide, with almost 2.5 million new cases representing 12.4% of all cancers. It was also the leading cause of cancer death, with around 1.8 million deaths accounting for 18.7% of all cancer deaths [7]. These statistics highlight the need for improved methods to better understand lung cancer biology and identify tissue patterns associated with patient outcome. The tumour microenvironment (TME), including the interaction between tumour cells and immune cells, plays an important role to understand the lung cancer biology for cancer progression, treatment response, and prognosis [45]. Multiplexed immunofluorescence (mIF) is an imaging technique that uses marker-specific antibodies with fluorescence signals to detect multiple proteins within the same tissue section. Each biological marker is captured as a separate image channel, allowing different cell types and their spatial patterns to be analyzed. mIF is therefore useful for studying the TME, as it preserves information about both protein expression and the spatial relationships between cells [23]. When combined with patient survival data, these tissue and cell patterns can be used to learn more about cancer biology and patient outcome. However, manually extracting meaningful information from mIF images is challenging, because these images contain complex and diverse spatial interactions within the TME [40]. In addition to that, image analysis can be affected by segmentation quality, staining variation, and tissue artifacts [42]. Therefore, AI-based methods are increasingly used to learn these complex and diverse representations directly from multichannel mIF images and to support the extraction of biological and prognostic insights [34]. The need for large annotated datasets is a major challenge in AI-based supervised deep learning methods. In mIF microscopy, annotations are difficult to obtain because they often require expert knowledge and detailed labelling at the cell or tissue level. The effort to develop AI-supported approaches that do not require extensive data annotation is rapidly increasing in digital pathology as an alternative to fully supervised methods. The self-supervised learning (SSL)-based approaches offer a promising way to reduce the burden from manual data annotation by learning representations directly from image data [25]. The learned representations from SSL can then be transferred to downstream tasks, such as classification, clustering, or survival prediction [47]. However, the majority of the current methods such as foundation models and pretrained backbones are mainly designed for RGB pathology images, and comparatively fewer methods have been developed for multichannel mIF imaging [31]. This motivates the need to explore the available representation learning strategies in order to find those which are best suited for multichannel mIF data. In this thesis, we explore two SSL-based approaches to extract meaningful representations from multichannel mIF images of lung cancer tissue. The extracted representations are then combined with patient-level survival data and evaluated in a downstream survival prediction task. Research Questions To address the motivation and challenges introduced above, this thesis aims to answer the following research questions: Can image representations learned in a self-supervised manner from multichannel Multiplexed immunofluorescence (mIF) microscopy images of cancer tissue provide prognostic information for lung cancer survival prediction? Which of the selected contemporary approaches achieves the best performance for survival prediction from multichannel mIF microscopy images? Aims The main aim of this thesis is to evaluate the performance of state-of-the-art SSL approaches to analyze multichannel mIF microscopy images in the task of predicting the survival of lung cancer patients. Using our own dataset [3] of lung cancer mIF tissue images combined with patient survival data, this project evaluates and compares contemporary representation learning approaches for this task. To address this aim, we divide the study into three main objectives. Identification and selection of two suitable contemporary approaches to be evaluated and compared. Implementation, evaluation, and comparison of the selected approaches on the given dataset with a baseline model. Preliminary exploration of possible improvements to the selected approaches.