Meme kanseri tespiti için sentetik mikrodalga görüntülerinin derin öğrenme odaklı segmentasyonu
Driven segmentation of synthetic microwave images for breast cancer detection
- Tez No: 949433
- Danışmanlar: PROF. DR. MEHMET ÇAYÖREN
- Tez Türü: Yüksek Lisans
- Konular: Elektrik ve Elektronik Mühendisliği, Mühendislik Bilimleri, Electrical and Electronics Engineering, Engineering Sciences
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2025
- Dil: Türkçe
- Üniversite: İstanbul Teknik Üniversitesi
- Enstitü: Lisansüstü Eğitim Enstitüsü
- Ana Bilim Dalı: Elektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Biyomedikal Mühendisliği Anabilim Dalı
- Sayfa Sayısı: 85
Özet
Günümüzde yapay zekâ tabanlı derin öğrenme modelleri, görüntü işleme ve bilgisayarlı görü alanlarında önemli gelişmelere öncülük etmektedir. Bu gelişmeler, özellikle tıbbi görüntüleme alanında hastalıkların erken teşhisi ve doğru tanılanması açısından büyük fayda sağlamaktadır. Kadınlarda en yaygın şekilde karşılaşılan kanser türlerinden biri meme kanseridir ve erken teşhis, tedavi başarısını doğrudan etkilemektedir. Bu nedenle, güvenli, taşınabilir ve iyonlaştırıcı radyasyon içermeyen bir yöntem olan mikrodalga görüntüleme, alternatif bir görüntüleme tekniği olarak dikkat çekmektedir. Ancak mikrodalga görüntülerin düşük çözünürlüklü ve gürültülü olması, doğrudan yorumlanmalarını güçleştirmektedir. Bu çalışmada, mikrodalga görüntülerde tümörlü dokuların daha net ve doğru şekilde belirlenebilmesi amacıyla derin öğrenme tabanlı bir segmentasyon yöntemi kullanılmıştır. Bu sayede, düşük kaliteli görüntülerden daha anlamlı bilgiler elde edilerek meme kanseri teşhisinde yardımcı olunması hedeflenmiştir. Bu çalışmada, bir adet tümör dokusu içeren sentetik meme yapısına ait toplam 1000 adet mikrodalga görüntü veri seti ve bu veri setine karşılık gelen segmentasyon görüntüleri girdi verisi olarak kullanılmıştır. Görüntülerin oluşturulabilmesi için 10 cm yarıçaplı 18 adet vivaldi anten içeren özel bir senaryo tasarlanmıştır. Bu senaryoda antenler, sentetik meme dokusunun etrafına yerleştirilerek farklı açılardan sinyal gönderip alacak şekilde konumlandırılmıştır. Bu sayede, tümörlü bölgenin mikrodalga sinyallerine verdiği tepkiler kaydedilmiş ve bu veriler kullanılarak dielektrik görüntüler oluşturulmuştur. Bu dielektrik görüntüler Reverse Time Migration (RTM) algoritması kullanılarak mikrodalga görüntülere çevrilmiştir. Elde edilen bu görüntüler, derin öğrenme tabanlı modellerin eğitimi ve testinde kullanılmıştır. Bu çalışmada iki farklı mimari kullanılmıştır: Autoencoder ve U-Net. Bu mimariler Python yazılım dili kullanılarak tasarlanmıştır. Bu iki mimari için üretilen 1000 adet görüntünün 600'ü eğitim, 200'ü validasyon ve 200'ü test için ayrılmıştır. Autoencoder modeli, gürültüyü azaltmak amacıyla girdiyi kodlayıp yeniden oluşturan bir yapıya sahiptir. Optimizasyon algoritması olarak Adam optimizasyonu kullanılmıştır. U-Net mimarisi ise gürültüyü temizlerken detayları daha iyi koruyabilmek amacıyla geliştirilmiştir. Eğitim sürecinde Binary Crossentropy kayıp fonksiyonu kullanılmıştır. Mimarilerin başarımı, Yapısal Benzerlik İndeksi (SSIM) metriği ile değerlendirilmiştir. SSIM değeri 0 ile 1 aralığındadır ve sonucun 1 olması orijinal görüntü ile tahmin edilen görüntünün birbiriyle tamamen aynı olduğu anlamına gelmektedir. Yapılan çalışma sonucunda autoencoder mimarisinin SSIM değeri 0,9419 iken, U-Net mimarisinin SSIM değeri 0,9627'dir. Bu sonuçlar doğrultusunda U-Net Mimarisi, autoencoder mimarisine göre %2,21 oranında mikrodalga görüntü segmentasyonunda daha başarılı olmuştur.
Özet (Çeviri)
Breast cancer remains one of the most prevalent and life-threatening diseases worldwide, continuing to be the leading cause of cancer-related mortality among women. Current global health statistics reveal that millions of new breast cancer cases are diagnosed annually, with a disproportionate concentration in low- and middle-income countries (LMICs). In these regions, disparities in healthcare infrastructure, socioeconomic challenges, and limited access to advanced diagnostic technologies substantially hinder early detection efforts. Early diagnosis of breast cancer is of paramount importance, as timely intervention significantly improves patient survival rates, reduces the need for aggressive treatments, mitigates treatment-related side effects, and ultimately decreases the economic burden on healthcare systems. Despite technological advancements, conventional breast imaging modalities such as mammography, ultrasound, and magnetic resonance imaging (MRI) still face notable limitations globally in terms of accessibility, cost, and diagnostic efficacy. Mammography, regarded as the clinical gold standard for breast cancer screening, involves ionizing radiation and poses potential health risks with frequent use. Moreover, its diagnostic sensitivity is considerably reduced in women with dense breast tissue, leading to false-negative results and delayed treatment. Radiation-free ultrasound imaging requires highly skilled operators, resulting in variability in diagnostic accuracy and limiting its widespread application. Nuclear imaging techniques such as positron emission tomography (PET) and single-photon emission computed tomography (SPECT) provide valuable dynamic and functional imaging data; however, their high costs, radiation exposure, and limited availability restrict their broader clinical use. MRI offers superior soft tissue contrast and spatial resolution but is associated with high operational costs, infrastructure demands, and limited accessibility, especially in LMICs. Collectively, these challenges underscore the urgent need for safe, cost-effective, scalable, and reliable novel imaging technologies for early tumor detection. Microwave Imaging (MWI) has recently emerged as a promising alternative modality that enables non-ionizing electromagnetic wave-based examination of the internal structural and functional properties of breast tissue. The fundamental principle of MWI relies on the dielectric property differences—specifically dielectric permittivity (∈_r) and electrical conductivity (σ)—between malignant tumors and normal breast tissue. This contrast provides functional imaging information complementary to the anatomical data obtained from conventional methods. Microwave imaging systems primarily consist of transmitting and receiving antennas, with the collected data computationally reconstructed into images. Additionally, MWI systems are inherently more affordable, portable, and easier to operate, making them highly suitable for resource-constrained clinical environments. Nevertheless, the clinical translation of MWI has been impeded by intrinsic limitations such as relatively low spatial resolution, sensitivity to noise and artifacts, and challenges in accurately localizing tumors within reconstructed images. In response to these challenges, the present thesis proposes a comprehensive deep learning framework aimed at enhancing the diagnostic performance and tumor localization accuracy of microwave breast imaging. Within this scope, synthetic breast tissue models were developed using a high-fidelity simulation platform based on the Reverse Time Migration (RTM) algorithm, which realistically models the propagation and scattering of electromagnetic waves in heterogeneous and homogeneous biological tissues. The RTM algorithm processes source and receiver wavefields bidirectionally and establishes correlations between them to generate an image. It consists of two steps involving the backpropagation of complex conjugate data into the imaging domain and computing the cross-correlation norm of the forward and backward propagated fields. Synthetic breast models were generated employing microwave difference imaging based on the total electric field, necessitating knowledge of the healthy breast state during imaging. The total electric field data required for microwave imaging were synthetically produced via numerical solutions of electromagnetic scattering problems. This study investigates deep learning-based segmentation of low-resolution microwave images generated from total electric field data. The dataset comprises 1000 synthetic transverse breast images, each containing a single tumor, initially created at a high resolution of 400×400 pixels. The use of synthetic data instead of physical breast phantoms provided precise control over critical parameters such as tissue heterogeneity, tumor size and location, and dielectric properties, allowing systematic and reproducible evaluation of model robustness under diverse conditions. The imaging configuration simulates a circular array of 18 Vivaldi antennas placed around the synthetic breast at a radius of 10 cm. The surrounding medium is modeled as a homogeneous environment with a relative permittivity of 10. Simulations incorporate properties of breast, fibroglandular, and tumor tissues. This arrangement ensures wide angular coverage and enhances sensitivity to tissue heterogeneity. To align with deep learning architectures and reduce computational load, images were resized to 64×64 pixels while preserving essential structural and contrast information necessary for accurate segmentation. It should be noted that microwave images are not inherently RGB; rather, single-channel microwave data were artificially converted into three-channel (64, 64, 3) pseudo-color images using assigned color maps. This transformation enables universal convolutional neural networks designed for RGB inputs to process the data. Both deep learning models used in this thesis accept input and produce output images of (64, 64, 3) dimensions, visually highlighting tumor regions to facilitate interpretation and improve suitability for segmentation tasks. Within this framework, two distinct deep learning architectures were proposed for tumor segmentation: an autoencoder and a U-Net model. Both architectures were implemented in Python using libraries such as Keras, TensorFlow, skimage, matplotlib, and numpy. Both models focus primarily on semantic segmentation of tumor tissues beyond mere noise removal or artifact suppression. The dataset of 1,000 images was split into 600 for training, 200 for validation, and 200 for testing on unseen samples. Training and testing were conducted in two separate phases; first, the systems were trained and the architectures saved, followed by testing using the trained models. The autoencoder comprises an encoder that compresses input images through multi-layer convolutional blocks incorporating batch normalization and LeakyReLU activation functions. Batch normalization accelerates training and stabilizes intermediate outputs, while LeakyReLU prevents“dead ReLU”problems by allowing small non-zero gradients for inactive neurons. Dropout regularization in deeper layers prevents overfitting. The encoder compresses fundamental spatial and structural features into latent representations, which the decoder reconstructs into enhanced images emphasizing tumor regions through transposed convolutions and additional convolutional operations. This architecture not only reconstructs image content but also improves contrast and structural integrity to support subsequent segmentation steps. However, since the autoencoder primarily aims at general image enhancement, its performance in precise tumor localization remains limited. In contrast, the U-Net model employed in this study is a supervised architecture specifically designed for biomedical image segmentation. U-Net features a symmetric encoder-decoder structure interconnected by skip connections that preserve high-resolution spatial information losslessly. The encoder extracts multi-scale features via convolution and max-pooling layers, while the decoder progressively upsamples these features to generate tumor probability maps. A sigmoid activation function in the final output layer produces binary segmentation masks with well-defined tumor boundaries. Unlike the autoencoder, U-Net is directly optimized for semantic segmentation using custom loss functions that consider both spatial accuracy and tumor region integrity. Consequently, it yields highly precise and clinically meaningful segmentation maps suitable for decision-making. Both models maintain structural symmetry between compression and expansion rates to ensure performance and stability during training. They were trained and validated on different subsets of the synthetic dataset with extensive data augmentation including rotation, reflection, and scaling to improve generalization and prevent overfitting. The Adam optimizer was used during training, with the autoencoder trained for 150 epochs and the U-Net for 50 epochs. Structural Similarity Index (SSIM) was employed as the primary evaluation metric to assess preservation of structural details and perceptual image quality during both training and testing phases. Comparative results demonstrate that the autoencoder effectively enhances overall image contrast and suppresses irrelevant details but remains limited in tumor localization due to its reconstruction-centric design. Conversely, the U-Net model significantly outperforms the autoencoder in both SSIM scores and segmentation accuracy, producing highly reliable tumor segmentation masks. In conclusion, this thesis presents a deep learning-based framework addressing fundamental limitations of microwave breast imaging. Utilizing high-fidelity synthetic datasets and robust neural network architectures, the proposed approach substantially improves tumor detection and localization capabilities. The non-invasive, radiation-free, and low-cost nature of microwave imaging renders this technology particularly promising for expanding early breast cancer screening in resource-limited settings worldwide. The proposed methods close current technological gaps by enabling more accurate tumor detection in low-resolution microwave images. Moreover, the successful integration of microwave imaging with deep learning not only overcomes existing technical challenges but also paves the way for more accessible and equitable breast cancer diagnostic methods globally. This approach reduces dependence on expensive and complex imaging equipment, holding transformative potential for screening protocols in disadvantaged regions where early detection remains critically insufficient. The innovations presented in this thesis represent a significant step toward reducing global health disparities, providing clinicians with advanced tools for timely diagnosis and personalized treatment planning. Ultimately, this work offers meaningful progress in the fight against breast cancer by combining technological innovation with intelligent healthcare solutions to enhance patient survival rates and quality of life.
Benzer Tezler
- Unet based segmentation in qualitative microwave imaging for breast cancer diagnosis
Meme kanseri tanisi için nitel mikrodalga görüntülemenin unet tabanli segmentasyonu
GİZEM ÜNER
Yüksek Lisans
İngilizce
2025
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
PROF. DR. MEHMET ÇAYÖREN
- Machine learning based augmentation of medical microwave imaging
Medikal mikrodalga görüntülemenin makine öğrenmesiyle iyileştirilmesi
MERVE KAPLAN ŞAFAK
Yüksek Lisans
İngilizce
2022
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
PROF. DR. MEHMET ÇAYÖREN
- Antenna design for breast cancer detection and machine learning approach for birth weight prediction
Meme kanseri tespiti için anten tasarımı ve doğum ağırlığı tahmini için makine öğrenmesi yaklaşımı
HALUK KIRKGÖZ
Yüksek Lisans
İngilizce
2024
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ ONUR KURT
- Makine öğrenmesi teknikleriyle mikrodalga meme kanserinin tespiti ve konumunun belirlenmesi
Detection and localization of microwave breast cancer by machine learning techniques
ELİF TÖLÜK
Yüksek Lisans
Türkçe
2022
MatematikYıldız Teknik ÜniversitesiMatematik Mühendisliği Ana Bilim Dalı
DOÇ. DR. BİROL ASLANYÜREK
- Development of sandwich type nucleic acid array platform for the detection of microRNAs in breast cancer
Meme kanseri tespiti için sandviç tabanlı nükleik asit dizi platformunun geliştirilmesi
SEREN ATILGAN
Yüksek Lisans
İngilizce
2014
Bilim ve TeknolojiOrta Doğu Teknik ÜniversitesiBiyoloji Ana Bilim Dalı
PROF. DR. HÜSEYİN AVNİ ÖKTEM