| Περίληψη: | Breast cancer is one of the leading health concerns, that affects millions of women worldwide. Breast
cancer detection plays a vital role in improving patient outcomes and survival rates. In recent years,
Computer-Aided Diagnosis (CAD) systems have emerged as powerful tools in medical imaging,
leveraging advanced deep learning algorithms to assist radiologists in accurately identifying potential
cancerous lesions. This thesis focuses on the implementation and evaluation of two CAD systems,
UNet and YOLOv5, for breast cancer detection.
The thesis begins by providing an overview of breast cancer, highlighting its prevalence and the
limitations of traditional diagnostic methods. It emphasizes the need for advanced Computer-Based
systems to augment radiologists’ expertise and improve diagnostic accuracy. UNet and YOLOv5,
both widely recognized in computer vision tasks, are implemented in order to aid breast cancer detection.
The technical aspects of UNet and YOLOv5, including their architectures and training procedures, are discussed in detail. The thesis addresses the challenges specific to medical imaging, such as
data preprocessing and augmentation techniques, and highlights the potential benefits of integrating
these models into the diagnostic workflow.
An extensive evaluation is conducted using a subset of digital mammograms of the INbreast
dataset. Metrics including Precision, Recall and F1-Score are used to measure their effectiveness.
Comparative analyses provide insights into the strengths and weaknesses of both networks. Specifically, the best results can be summarized on average for the UNet as Precision: 85.87%, Recall:
85.10% and F1-Score: 84.25%. For the UNet CLAHE as Precision: 89.73%, Recall: 84.12% and F1-
Score: 85.13%. Lastly for the YOLOv5 as Precision: 92.30%, Recall: 90.60% and F1-Score: 91.20%.
Recommendations and further improvements, such as model optimization are also included.
The conclusions drawn from this thesis highlight the potential of UNet and YOLOv5 in breast
cancer detection. The implemented CAD systems demonstrate the ability to accurately detect cancerous lesions in medical images, aiding radiologists in making timely and informed diagnoses. The
findings emphasize the importance of comprehensive and diverse datasets, as well as collaboration
between medical professionals and computer scientists, to optimize CAD systems and drive advancements in breast cancer detection.
Overall, this thesis contributes to the research on CAD systems, and their potential to improve
the accuracy and efficiency of breast cancer diagnosis. Continued research in this field holds promise
for further advancements, ultimately leading to enhanced patient outcomes and a significant reduction in breast cancer-related morbidity and mortality.
|
|---|
