Πτυχιακές εργασίες

Μόνιμο URI για αυτήν τη συλλογήhttps://pyxida.aueb.gr/handle/123456789/11719

Περιήγηση

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω

Τώρα δείχνει 1 - 2 από 2

From pen to prediction: handwriting-based Alzheimer’s detection
(2025-10-29) Boumpi, Maria; Μπουμπή, Μαρία; Pavlopoulos, John
Detecting Alzheimer’s disease (AD) at an early stage is critical for planning effective treatments, supporting patients and their families, and slowing the progression of symptoms. Conventional practices, such as neuroimaging and biomarker assessment, tend to be costly and impractical for broad screening applications. This study tries to fill the gaps from previous studies and explore the ability of handwriting analysis as a non-invasive, accessible, and affordable way to identify AD. Our approach combines two perspectives, measurable features from pen movement patterns and handwritten images. The analysis integrated two datasets, the DARWIN dataset, which offers a range of tabular and image data from handwriting samples from multiple tasks, and the Alzheimer’s Disease Dataset (ADD), a detailed clinical dataset with demographic, medical, and cognitive assessment data. A series of classification approaches was applied, classic machine learning models (Random Forest, SVM, XGBoost) on tabular data, a deep learning-based Swin Transformer for images, and a multimodal classifier fusing the two modalities. The results confirm that the handwriting-based features are individually strong enough to be used as a diagnostic tool, as Random Forest achieved 83.03% ± 1.18 on DARWIN tabular data and XGBoost achieved 83.53% ± 3.44 on the ADD dataset. The Swin Transformer also managed to reach a consistent performance on handwriting images (80.02% ± 0.87) and was able to capture delicate motor and spatial anomalies that suggest cognitive decline. The best overall performance was achieved with a late fusion model that combined both modalities, achieving 89.15% ± 1.73 Accuracy. This highlights how combining visual and structured data can capture both neuromotor and cognitive impairment. Ablation studies looked at the effects of sequence and order on handwriting tasks and the bounds of joint training in data-constrained fusions. Studies have demonstrated that modular fusion provides more consistent and interpretable outcomes than any other method available in small-sample situations. In summary, handwriting, represented by measurable features from pen movement patterns and raw handwriting images, is simple, inexpensive, and reliable for detecting AD. It surpasses clinical performance while enabling home and clinical use.
Enhanced biomedical image tagging
(2024-09-15) Chatzipapadopoulou, Anna; Χατζηπαπαδοπούλου, Άννα; Androutsopoulos, Ion; Pavlopoulos, John
Medical image classification is a complex and fascinating task that has significantly benefited from advancements in deep learning techniques. This thesis focuses on the specific challenge of multi-class multi-label medical image classification, also known as medical image tagging. The objective is to accurately assign relevant medical terms (concepts/tags) to images, which describe potential findings and ultimately assist clinicians in the diagnostic process. Medical image tagging may involve analyzing images from various modalities, such as X-rays, MRIs, CT scans, and ultra sonographies, to identify and categorize different pathological conditions. To address the task of medical image tagging, we develop and evaluate various deep learning models that encode the images and combine them with both classification-based and retrieval-based approaches. Ultimately, our goal is to develop a robust and reliable system for medical image tagging that not only provides clinicians with supportive information but also assists them in diagnosing diseases more accurately and efficiently.

Περιήγηση

Πλοήγηση Πτυχιακές εργασίες ανά Θέμα "Deep learning"