The Dark Side of the Eye: Towards Robust Deep Learning Models for Uveal Melanoma Screening

UM (Uveal Melanoma) is a rare but highly aggressive cancer, characterized by high metastatic risk and a poor prognosis if not treated in time. Early detection is challenging, as this lesion presents overlapping features with benign ones, which can lead to misdiagnosis. A late diagnosis often results in invasive treatments, which deeply affect both the patient's quality of life and the healthcare system. For this reason, improving the efficiency and reliability of screening procedures is crucial. Typically, a diagnosis of UM is made through manual ophthalmic image analysis. This method allows for non-invasive exploration of ocular structures with high accuracy and precision. However, the procedure is time-consuming, and accurate interpretation requires significant expertise, which is often scarce due to the limited number of specialized ocular oncologists. As a result, many patients are initially evaluated by general ophthalmologists who may not have the expertise needed to distinguish subtle differences between early-stage melanomas and benign lesions, leading to delayed diagnoses. Consequently, diagnoses are often made when the tumor has already reached an advanced stage. Therefore, the development of automated solutions for screening is of paramount importance to speed up the analysis process and reduce inter-operator variability. In this context, Deep Learning (DL) presents a promising avenue because it can handle large quantities of high-dimensional data, such as images. However, two main challenges need to be addressed. First, the state-of-the-art research related to UM screening is fragmented. Few solutions are available in the literature, with most studies focusing on analyzing tissue images acquired through biopsies. While these studies are valuable for precise analysis of tumor subtyping and treatment planning, they are not suitable for screening purposes due to the invasive nature of a biopsy. As a result, a standardized baseline for automated screening is still missing. Second, there is an inherent issue with data scarcity, stemming from the rarity of this disease, which impedes the development of efficient solutions. This thesis aims to address these challenges by establishing a baseline for automated ocular lesion classification using common image modalities at the macroscale, while also exploring strategies to enhance robustness in the face of data scarcity. Specifically, this work draws inspiration from recent studies suggesting that reframing the classification task into the more complex task of segmentation can help achieve greater robustness in conditions of data scarcity. However, this hypothesis has not been systematically validated in real-world clinical scenarios. Additionally, most previous research has primarily focused on semantic segmentation, leaving the question open as to whether other paradigms, such as detection, could offer similar improvements in robustness. To fill these gaps, this thesis evaluates various model architectures for differentiating between different ocular lesions, assessing their performance under conditions of data scarcity. The goal is to identify which models are more robust when trained with limited data. In the next phase, this thesis builds upon the findings of the investigation to propose a processing strategy followed by a training procedure, aimed at optimizing model performance and improving clinical applicability. The investigation carried out in this thesis led to the definition of an initial strategy to enhance model robustness for automated screening of UM, providing a foundation for future studies and advancements in this field.

Author dc:creator

• Virginia Tasso (24399065)

• Virginia Tasso

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.32991959.v1

OAI identifier oai:identifier

oai:figshare.com:article/32991959