3B Medikal Görüntü İşleme İçin Derin Öğrenme Model Mimarisinin Geliştirmesi ve Analizi
No Thumbnail Available
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Open Access Color
OpenAIRE Downloads
OpenAIRE Views
Abstract
Günümüzde medikal görüntü segmentasyonuna yönelik geliştirilen derin öğrenme modelleri, yüksek doğruluk sunmalarına rağmen; aşırı hesaplama maliyeti, karmaşık yapılar ve donanım bağımlılığı nedeniyle pratik kullanımda çeşitli sınırlılıklar barın-dırmaktadır. Bu doğrultuda, kullanıcı dostu, düşük donanım gereksinimiyle çalışabi-len, sade ancak derin yapıda, sınırlı veri setlerinde de etkili sonuçlar verebilen, genellenebilir ve güçlü mimarilere duyulan ihtiyaç giderek artmaktadır. Bu tezde, herhangi bir fine-tuning veya dışsal optimizasyona ( pruning, quantization, attention vb.) ihtiyaç duymadan, yalnızca yapısal mimari iyileştirmelerle yüksek doğruluk elde eden donanım dostu bir 3B CNN modeli geliştirilmiştir. Model mimarisi kapsamlı biçimde ele alınmış; katman derinliği, filtre boyutu, kanal sayısı, aktivasyon ve normalizasyon sıralaması gibi birçok parametre sistematik olarak analiz edilmiştir. Farklı çekirdek boyutlarına sahip konvolüsyon filtreleri hem paralel yollarla aynı blok içinde, hem de ardışık katmanlar arasında dağıtılarak farklı mimari konfigürasyonlarla yapılandırılmıştır. Bu yapılarda tek ve çok katmanlı, simetrik ve asimetrik tasarımlar denenmiştir. Ayrıca model tasarımı sürecinde NAS (Neural Architecture Search) yöntemi uygulanmış; elde edilen mimari varyantlar performans açısından değerlendirilmiştir. Geliştirilen model, klasik U-Net'e kıyasla eğitim süresini 2.5 ila 10 kat arasında kısaltmış, FLOPs değerini yaklaşık yarı yarıya düşürmüş ve benzer Dice Benzerlik Katsayısı (DSC) ile segmentasyon doğruluğunu korumayı başarmıştır. Ayrıca yapılan analizlerde, FLOPs'un gerçek zamanlı performansı belirlemede tek başına yeterli bir ölçüt olmadığı ortaya konmuştur. Bu tez kapsamında yürütülen çalışmalar, yalnızca mimari düzeyde gerçekleştirilen iyileştirmelerle yüksek doğruluk ve donanım verimliliğine ulaşılabileceğini göstermekte; geliştirilen yapının sade fakat derin mimarisi-yle genellenebilirliği, sınırlı veri setlerinde başarımı ve hangi mimari parametrelerin modele belirgin katkı sağladığı detaylı biçimde ortaya konmuştur.
Deep learning based models developed for medical image segmentation achieve high accuracy, yet their practical implementation remains limited due to excessive computational cost, architectural complexity, and dependency on hardware resources. Accordingly, there is an increasing demand for models that are structurally compact, hardware efficient, sufficiently deep, generalizable, and capable of performing effectively even with limited datasets. This thesis presents a 3D convolutional neural network architecture that achieves high segmentation accuracy through architectural modifications alone, without requiring any external optimization techniques. The architecture was thoroughly investigated, and key parameters such as layer, kernel size, number of channels, and the ordering of activation and normalization were systematically analyzed. Convolutional kernels with varying receptive field sizes were employed in different configurations, including parallel branches within the same block and sequential layers across the architecture. The design process involved comprehensive experimentation using neural architecture search through which various architectural configurations were evaluated. Compared to the classical U-Net model, the proposed architecture reduced training time by a factor of 2.5 to 10, halved the number of floating point operations, and maintained comparable segmentation accuracy in terms of the Dice Similarity Coefficient.Furthermore, it was observed that although FLOPs is widely used as a computational cost metric, it does not directly correlate with actual inference time. The findings of this thesis demonstrate that it is possible to achieve high segmentation performance and hardware efficiency through architectural design alone, and offer a detailed investigation into the structural parameters that most significantly affect performance in compact and generalizable CNN models.
Deep learning based models developed for medical image segmentation achieve high accuracy, yet their practical implementation remains limited due to excessive computational cost, architectural complexity, and dependency on hardware resources. Accordingly, there is an increasing demand for models that are structurally compact, hardware efficient, sufficiently deep, generalizable, and capable of performing effectively even with limited datasets. This thesis presents a 3D convolutional neural network architecture that achieves high segmentation accuracy through architectural modifications alone, without requiring any external optimization techniques. The architecture was thoroughly investigated, and key parameters such as layer, kernel size, number of channels, and the ordering of activation and normalization were systematically analyzed. Convolutional kernels with varying receptive field sizes were employed in different configurations, including parallel branches within the same block and sequential layers across the architecture. The design process involved comprehensive experimentation using neural architecture search through which various architectural configurations were evaluated. Compared to the classical U-Net model, the proposed architecture reduced training time by a factor of 2.5 to 10, halved the number of floating point operations, and maintained comparable segmentation accuracy in terms of the Dice Similarity Coefficient.Furthermore, it was observed that although FLOPs is widely used as a computational cost metric, it does not directly correlate with actual inference time. The findings of this thesis demonstrate that it is possible to achieve high segmentation performance and hardware efficiency through architectural design alone, and offer a detailed investigation into the structural parameters that most significantly affect performance in compact and generalizable CNN models.
Description
Keywords
Elektrik ve Elektronik Mühendisliği, Electrical and Electronics Engineering
Turkish CoHE Thesis Center URL
Fields of Science
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
115
Collections
Google Scholar™
Sustainable Development Goals
7
AFFORDABLE AND CLEAN ENERGY
