Medical Image Classification with Limited Data
Name
Glib Manaiev
Abstract
Advancements in computational methods have greatly influenced medical
imaging, facilitating the development of advanced diagnostic tools. One of the many tasks in this field is classifying images to determine whether they contain disease. This task is challenging because of the scarcity of annotated medical data, as it is harder to annotate because it requires an expert.
Lately, the problem of limited annotations has often been addressed by a group
of learning approaches that utilize unannotated data, known as unsupervised learning. Typically, models are pretrained on an artificial task that exploits the properties of images, rather than their annotations, and then fine-tuned on annotated data. Despite the recent success of these methods, they remain minimally explored in the field of medical imaging, particularly in medical image classification.
This thesis investigates the effectiveness of various unsupervised pretraining ap-
proaches in enhancing the classification of medical images, specifically focusing on kidney tumor classification from CT (computed tomography) scans, which represents a distinct challenge within medical image classification. In our experiments, these methods do not significantly improve model performance, but offer insights into the limitations and possibilities of unsupervised learning in this area. Contrary to prior expectations about the transformative impact of unsupervised pretraining, the benefits appear dependent on specific contexts and tasks. This work illustrates the complexity of enhancing model performance in this field, emphasizing the need for a comprehensive approach to tackling these challenges.
imaging, facilitating the development of advanced diagnostic tools. One of the many tasks in this field is classifying images to determine whether they contain disease. This task is challenging because of the scarcity of annotated medical data, as it is harder to annotate because it requires an expert.
Lately, the problem of limited annotations has often been addressed by a group
of learning approaches that utilize unannotated data, known as unsupervised learning. Typically, models are pretrained on an artificial task that exploits the properties of images, rather than their annotations, and then fine-tuned on annotated data. Despite the recent success of these methods, they remain minimally explored in the field of medical imaging, particularly in medical image classification.
This thesis investigates the effectiveness of various unsupervised pretraining ap-
proaches in enhancing the classification of medical images, specifically focusing on kidney tumor classification from CT (computed tomography) scans, which represents a distinct challenge within medical image classification. In our experiments, these methods do not significantly improve model performance, but offer insights into the limitations and possibilities of unsupervised learning in this area. Contrary to prior expectations about the transformative impact of unsupervised pretraining, the benefits appear dependent on specific contexts and tasks. This work illustrates the complexity of enhancing model performance in this field, emphasizing the need for a comprehensive approach to tackling these challenges.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Dmytro Fishman, Joonas Ariva
Defence year
2024