University of Peloponesse
Μελέτη και ανάπτυξη συστήματος αυτόματης ανίχνευσης και αναγνώρισης ύποπτων καρκινικών περιοχών σε ακτινογραφίες μαστογραφίας
Abstract
dc:descriptionBreast cancer remains a significant global health issue, ranking as the second most prevalent cancer worldwide and the most frequently diagnosed cancer among women. The paramount importance of early, automated, and precise detection cannot be overstated, as it is intrinsically linked to improved patient prognosis and more effective treatment strategies. Advanced medical imaging plays a pivotal role in the detection, diagnosis, and management of breast cancer. The primary imaging modalities employed include mammography, encompassing Full-Field Digital Mammography (FFDM) and Digital Breast Tomosynthesis (DBT), breast ultrasound (US), and Magnetic Resonance Imaging (MRI). Each of these modalities possesses distinct advantages and limitations in the context of breast cancer screening and diagnosis. Mammography, particularly FFDM, is a cornerstone of breast cancer screening, primarily due to its efficacy in identifying microcalcifications, which are often early indicators of malignancy. Digital mammography offers advantages in image storage and sharing, while DBT, or 3D mammography, enhances visualization by creating cross-sectional images of the breast, thereby reducing tissue overlap and potentially improving cancer detection rates, especially in women with dense breast tissue. The continued reliance on FFDM, even with its recognized limitations such as observer variability and the aforementioned workload and diagnostic errors, highlights a significant opportunity for artificial intelligence (AI) driven solutions, like Generative Adversarial Networks (GANs), to make a substantial contribution. Such contributions could extend beyond merely aiding in cancer detection to potentially optimizing screening workflows, standardizing image interpretation aspects, or enhancing the training of radiologists. Traditional AI and Computer-Aided Detection (CAD) systems, while beneficial, have their own constraints, including significant computational overhead or a propensity for errors, particularly false positives. Compounding these issues is the common scenario in medical imaging where datasets are relatively small, imbalanced (e.g., far more normal exams than cancerous ones), and subject to stringent privacy regulations that limit sharing and aggregation. GANs present a promising avenue to address these multifaceted problems by offering capabilities such as the generation of diverse and realistic synthetic training data to improve the robustness and generalization of diagnostic models, the synthesis of images to simulate rare pathological conditions or specific tissue characteristics (e.g., varying breast densities), the enhancement of image quality through denoising or super-resolution, and the potential creation of images for training and research purposes without exposing real patient data, thereby mitigating privacy concerns. Thus, the increasing adoption of GANs in breast imaging can be understood as a targeted and strategic response to these well-recognized and critical challenges within the field. Moreover, the specific ways in which GANs are applied are often meticulously tailored to address the inherent weaknesses or to amplify the strengths of particular breast imaging modalities. For instance, mammography's effectiveness is known to be reduced in dense breast tissue where lesions can be obscured; consequently, GANs are being developed to synthesize realistic high-density mammograms for training AI to perform better in such conditions, or to directly highlight lesions embedded within dense tissue. Early detection of breast abnormalities using mammography is critical to improving prognosis and survival. However, dense breast tissue can severely affect the diagnostic accuracy of mammography as it can mask underlying malignancies, making early detection more difficult. To tackle this urgent problem, our research presents a novel method to convert breast density ROIs in mammograms through generative adversarial networks (GANs), particularly CycleGAN and its variant GAN Hopper. The objective is to diminish the masking influence of dense tissue, thus enhancing the detection of hidden malignancies. Our method employs unsupervised image-to-image translation to convert mammographic images of high breast density (ACR-D) to low density (ACR-A) and identify hidden lesions while maintaining the original diagnostic characteristics. We utilized this approach on various mammogram datasets and showcased its efficacy in diverse environments. Experimental findings indicate that this approach has achieved meaningful advancements in identifying possible tumors concealed by dense breast tissue. Our research presents an innovative mammogram evaluation approach utilizing advanced machine learning methods, aimed at enhancing the diagnostic precision of dense breast tissue and anticipated to enhance the identification and management of early breast cancer.
Degree
thesis:*- Grantor dc:publisher
- University of Peloponesse
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Anyfantis, Dionysios
- Ανυφαντής, Διονύσιος
Subjects
dc:subject × 22- Σύστημα υποβοηθούμενης διάγνωσης
- Μαστογραφία
- Διάγνωση καρκίνου μαστού
- Βαθιά μάθηση
- Παραγωγικά ανταγωνιστικά δίκτυα
- Μετασχηματισμός εικόνας σε εικόνα
- Mαστική πυκνότητα
- Computer aided diagnosis
- Μammography
- Βreast cancer diagnosis
- Generative Adversarial Networks (GANs)
- Image-to-image translation
- Breast density
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Τεχνολογία μέσων
- Βιοϊατρική Μηχανική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Media Technology
- Medical Engineering
- Biomedical Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60593