Beyin tümörü segmentasyonu için yeni bir üç boyutlu U-Net tabanlı derin öğrenme mimarisi
A novel three-dimensional U-Net based deep learning architecture for brain tumor segmentation
- Tez No: 1018551
- Danışmanlar: PROF. DR. DEVRİM AKGÜN
- Tez Türü: Doktora
- Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2026
- Dil: Türkçe
- Üniversite: Sakarya Üniversitesi
- Enstitü: Fen Bilimleri Enstitüsü
- Ana Bilim Dalı: Bilgisayar Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Bilgisayar Mühendisliği Bilim Dalı
- Sayfa Sayısı: Belirtilmemiş.
Özet
Beyin tümörleri dünya genelinde en yüksek ölüm oranına sahip, yaşamı tehdit eden, erken dönemde teşhis ve tedavi edilmesi gereken vücut kitleleridir. Sağlıklı dokular ve tümörlü yapıların ayrıldığı manuel segmentasyonun sıkıcı, zaman alıcı ve uzman değerlendirmelerinde değişiklik göstermesi sebebiyle tam otomatik segmentasyon bir çözüm olarak görülmüştür. Manyetik Rezonans Görüntüleme (MRI) ile beyin tümörlerinin otomatik olarak segmentasyonu, teşhis ve tedavi planları geliştirmek için birincil tıbbi görüntüleme yöntemidir. Bununla birlikte, tümör konumu ve boyutundaki hastalar arası değişkenlik, segmentasyon için zor bir sorundur. Son zamanlarda özellikleri otomatik olarak öğrenen derin öğrenme modellerinin yükselişiyle tıbbi alanda kullanımı artmıştır. Bu tez çalışması, beyin tümörü segmentasyonu için tıbbi görüntü işleme alanında en popüler olan U-Net mimarisine dayalı çok ölçekli üç boyutlu hafif ve kademeli otomatik bir model (LCBTS-Net) sunmaktadır. Hafif ve kademeli mimari, aşırı uyum riskini en aza indirmeye yardımcı olur ve daha kararlı ve güvenilir tahminler sunar. Her MRI dizisi farklı anatomik ve patolojik özellikleri vurguladığından, model bu kanalları mevcut tüm bilgileri kullanmak için birleştirir ve ağ kodlayıcısına besler. Önerilen model, beyin tümörü alt yapılarının hiyerarşik yapısını kabul eden ve bunları kaba ayrıntıdan ince ayrıntıya kadar bölgelere ayıran çok adımlı kademeli bir tahmin yöntemi kullanır. Segmentasyon tahmini sırasında, tüm tümör (WT) bölgesinin sonucu, daha ince segmentasyon sürecine rehberlik etmek için tümör çekirdeği (TC) ve aktif tümör bölge (ET) tahmini için önsel bilgi olarak kullanılır. Böylece daha odaklı ve hassas tahmin gerçekleştirilir. Enerji sürdürülebilirliği hedefleri doğrultusunda geliştirilen bu hafif model, her tümör bölgesi için yalnızca 1.58 milyon parametre ve 247.09 GFlops içerdiğinden, özellikle sınırlı ortamlardaki gerçek dünya klinik uygulamaları için kullanışlı bir çözümdür. Model, Beyin Tümörü Görüntü Segmentasyon Mücadelesi (BRATS) 2020 bölünmüş test kümesinde WT, TC ve ET için sırasıyla 0.9285, 0.8871, 0.8694 Dice puanları ve 4.26, 6.46 ve 4.55 Hausdorff95 mesafeleriyle umut verici sonuçlar elde etmiştir. Modelin genelleştirilebilirliğini zenginleştirmek ve sabit bir test setinin neden olabileceği değerlendirme yanlılığını azaltmak amacıyla BRATS 2020 eğitim veri seti üzerinde 5 katlı çapraz doğrulama yapılmıştır. Doğrulama sonuçları modelin aşırı öğrenme veya yetersiz öğrenme yapmadan, yüksek genelleme yeteneği kazandığını ispatlamaktadır. Ayrıca, BRATS 2021 bölünmüş test verisinde WT, TC ve ET için sırasıyla 0.9472, 0.9238 0.9194 gibi rekabetçi Dice sonuçları elde etmiştir. Bu bulgular, önerilen yaklaşımın beyin tümörü segmentasyonu performansını mevcut en son teknoloji tekniklerine kıyasla önemli ölçüde iyileştirdiğini göstermektedir. Bu bağlamda LCBTS-Net, hafif ve modüler tasarımı sayesinde farklı klinik segmentasyon problemlerine uyarlanabilen genel amaçlı, temel bir mimari olarak değerlendirilebilir.
Özet (Çeviri)
Brain tumors are critical neurological disorders caused by abnormal and uncontrolled cell growth in brain tissue. Accurate detection and precise identification of tumor regions are crucial for diagnosis, treatment planning, surgical guidance, and treatment follow-up. MRI is the most commonly used method for imaging brain tumors due to its superior soft tissue contrast and multimodal imaging capabilities. However, manual segmentation of tumor regions from MRI scans is time-consuming, tedious, and highly dependent on expert experience. Consequently, automated brain tumor segmentation has become a significant research topic in medical image analysis and computer-aided diagnostic systems. Fully automated brain tumor segmentation is divided into two categories: traditional approaches and artificial intelligence based approaches. In traditional techniques, features (texture, shape, location, density, etc.) are extracted manually and fed to machine learning algorithms. Due to the complexity of brain structure, these methods, based on human expertise, have failed to generalize. Recently, with the rise of deep learning models that automatically learn features, their use in the medical field has increased. Deep learning models enable the learning of local and global information of images and allow for accurate and stable segmentation. The most popular deep learning model in medical image segmentation is U-Net. It achieves good performance even with a small number of training examples. Many of the techniques used in brain tumor segmentation have been created by modifying the U-Net network and have demonstrated high performance. BRATS competitions have played a significant role in standardizing the segmentation of tumor subregions. The BRATS competition has enabled the development and comparison of automated algorithms specifically designed to cope with the challenges of gliomas and other brain tumors. Automated segmentation of brain tumors via MRI is vital for diagnosis and the development of personalized treatment plans. However, inter-patient variability in tumor location and size makes automated MRI image segmentation a challenging issue. This thesis presents a multi-scale automated 3D lightweight and cascaded brain tumor segmentation model (LCBTS-Net) based on the U-Net architecture to accurately separate tumor regions from healthy tissues. The lightweight and cascade architecture helps minimize the risk of overfitting and provides more stable and reliable predictions. Since each MRI sequence highlights different anatomical and pathological features, the model combines these channels and feeds them to the network encoder to utilize all available information. The proposed model employs a multi-step cascade prediction method that acknowledges the hierarchical structure of brain tumor substructures and divides them into regions from coarse to fine detail. During segmentation prediction, the result of the WT region is used as prior information for TC and ET region prediction to guide the finer segmentation process. Thus, more focused and precise prediction is achieved. Developed in line with hardware constraints and energy sustainability goals, this lightweight model contains only 1.58 million parameters and 247.09 GFlops per tumor region, making it a useful solution for real-world clinical applications, especially in confined environments. The model achieved promising results in the BRATS 2020 dataset, with Dice scores of 0.9285, 0.8871, and 0.8694 for WT, TC, and ET, respectively, and Hausdorff95 distances of 4.26, 6.46, and 4.55. To enhance the model's generalizability, provide equal training opportunities for all patients, and reduce assessment bias that might arise from a fixed test set, a 5-fold cross-validation was performed on the BRATS 2020 training dataset. The results showed that the mean Dice scores of the tumor regions were quite close to each other. This demonstrates that the model achieved high generalization ability without over-learning or under-learning. Furthermore, it obtained competitive Dice results in the BRATS 2021 dataset, such as 0.9472 for WT, 0.9238 for TC, and 0.9194 for ET. The model achieved better qualitative and quantitative results compared to many other methods. High Dice values were obtained, particularly in WT segmentation, while TC and ET achieved comparable accuracy levels to architectures in the literature. These findings demonstrate that the proposed approach significantly improves the segmentation performance of brain tumor regions compared to state of the art techniques. One of the key strengths of this architecture is its ability to decompose the segmentation task into a cascade of binary subtasks. First, the WT is identified, then the smaller and more complex TC and ET regions are segmented. This method effectively reduces inter-class ambiguity and increases focus on less-represented anatomical regions. Identifying the WT first, then its associated TC and ET regions, narrows the search space. This helps reduce false positives and allows for more consistent tumor boundary determination. Consequently, it enables more accurate segmentation of challenging areas, such as TC and ET, compared to traditional multi-class 3D U-Net models. Furthermore, the modular architecture provides an adaptable framework for different datasets or organ segmentation problems. The model's inference efficiency provides a significant advantage for practical applications. Unlike many high-accuracy 3D architectures, this proposed design requires less graphics processing unit memory and computation time while still providing satisfactory segmentation performance. This makes it suitable for use in clinical workflows and edge devices. However, some limiting factors may affect the model's performance. The number of filters in the encoder convolution layers remains fixed at 64 to avoid increasing computational cost. Although the model is lightweight during inference, it incorporates three U-Net models, resulting in multi-model training. Another limitation of the proposed cascaded model is its susceptibility to segmentation errors in previous stages. Misclassifications in the WT stage can carry over to the TC and ET stages, leading to error accumulation, especially in anatomically small or ambiguous regions such as ET. Furthermore, significant fluctuations in TC and ET Dice scores were observed between validation stages. The cause of the fluctuation has been attributed to data heterogeneity (particularly the imbalance in the distribution of low- and high-grade gliomas) and to the difficulty of segmenting small objects. While this reflects clinical reality, it complicates consistent model performance. Furthermore, rare instances of model failure have been identified in cases with low resolution or unclear boundaries. Such situations are among the common causes of failure for medical image segmentation models. Nevertheless, when overall performance is evaluated, the proposed architecture has sufficient representational power despite its lightweight structure. Several research directions are suggested for addressing these limitations in future studies. Firstly, the use of uncertainty modeling or attention mechanisms within a hierarchical structure could enable the network to make more cautious predictions in regions with low confidence levels. In particular, integrating channel- and spatial-attention modules could help the model focus more on clinically significant ET and TC regions. Similarly, the addition of residual or dense connection blocks could help to transmit multi-scale features and reduce information loss more robustly. Combining transformer-based hybrid architectures, which have become widespread in recent years, with a hierarchical segmentation structure also holds significant research potential. From a data perspective, expanding data diversity is critical to improving model generalization capability. Methods such as integrating multicenter MRI datasets, using field-adaptation techniques, and generating generative adversial network-based synthetic MRIs can enhance the model's robustness across different imaging conditions. Furthermore, sampling strategies or loss functions designed to reduce class imbalance can improve segmentation accuracy, particularly for small ET regions. The potential for application of the proposed architecture is not limited solely to brain tumor segmentation. Similar hierarchical approaches can be practical in other medical imaging problems involving multi-scale and hierarchical structures, such as liver and lesion segmentation, kidney tumor analysis, or multi-organ segmentation. In this context, LCBTS-Net can be considered a general-purpose, basic architecture that can be adapted to different clinical segmentation problems thanks to its lightweight, modular design.
Benzer Tezler
- Vascular segmentation of brain MR angiography images using convolutional neural networks
Evrişimsel sinir ağları kullanarak beyin MR anjiyografi görüntülerinin vasküler segmentasyonu
YUSUF HÜSEYİN ŞAHİN
Yüksek Lisans
İngilizce
2017
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
PROF. GÖZDE ÜNAL
- Arttırılmış gerçeklik (AR) tabanlı nöronavigasyon tekniği geliştirilmesi ve bu tekniğin optik izleme tabanlı nöronavigasyon tekniği ile karşılaştırılması
3D printed marker-based augmented reality neuronavigation: new neuronavigation technique and comparison with optic tracking system neuronavigation
GÖRKEM YAVAŞ
Tıpta Uzmanlık
Türkçe
2021
NöroşirürjiEge ÜniversitesiBeyin ve Sinir Cerrahisi Ana Bilim Dalı
PROF. DR. MEHMET SEDAT ÇAĞLI
- Image segmentation and classification based on CNN model to detect brain tumor
Başlık çevirisi yok
NOOR S. SALEH
Yüksek Lisans
İngilizce
2022
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Okan ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
PROF. DR. SHADİ M S HİLLES
- Deep learning medical applications
Derin öğrenme tıbbi uygulamaları
ALI AMER SHYAA SHYAA
Yüksek Lisans
İngilizce
2020
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolAltınbaş ÜniversitesiElektrik ve Bilgisayar Mühendisliği Ana Bilim Dalı
Prof. Dr. GALİP CANSEVER
- Detection and classification of brain tumors in MRI images using deep convolutional neural network
Başlık çevirisi yok
HUSSEIN ALISMAEELI
Yüksek Lisans
İngilizce
2022
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolAltınbaş ÜniversitesiBilişim Teknolojileri Ana Bilim Dalı
PROF. DR. OSMAN NURİ UÇAN