Renal hücreli karsinomun otomatik derece sınıflandırması için U-net tabanlı derin öğrenme ağı
U-net based deep learning network for automatic grade classification of renal cell carcinoma
- Tez No: 948937
- Danışmanlar: PROF. DR. MÜRVET KIRCI
- Tez Türü: Yüksek Lisans
- Konular: Biyomühendislik, Bioengineering
- 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 Bilim Dalı
- Sayfa Sayısı: 83
Özet
Bu tez çalışmasında, RHK'nın WHO/ISUP derecelendirme sistemine uygun şekilde otomatik sınıflandırılması amacıyla çok görevli bir derin öğrenme modeli geliştirilmiştir. Önerilen model, U-Net mimarisinin enkoder yapısını temel alarak hem segmentasyon hem de sınıflandırma görevini eş zamanlı olarak gerçekleştirmektedir. Modelin dikkat mekanizması, hem Squeeze-and-Excitation (SE) blokları hem de çekirdek maskelerinden türetilmiş kanal dikkat yapısıyla güçlendirilmiştir. Böylece ağ, görüntüdeki hücre çekirdeklerine odaklanarak sınıflandırma kararlarını WHO/ISUP derecelendirme sistemine uygun düşecek şekilde bu bölgelerden gelen bilgiyi önceliklendirerek vermektedir. Model, KMC (Kasturba Medical College) veri seti kullanılarak eğitilmiş ve değerlendirilmiştir. Veri seti, 722 adet orijinal boyutu 1920x1440 olan görüntüden elde edilmiş birbiriyle örtüşmeyen 224x224 boyutlu yamaların yatay ve dikey yansıtma işlemleriyle arttırılmış, toplam 4077 adet etiketli (Derece 0–4) histopatoloji yamasından oluşmaktadır. Eğitimden önce 3 kanallı görüntüler her bir kanal için ayrı ayrı hesaplanmış ortalama ve standart sapma değerleriyle normalize edilmiştir. Eğitim sürecinde sınıflandırma için çapraz entropi kaybı ve hücre çekirdeği segmentasyonu işlemi için ikili çapraz entropi kayıp fonksiyonu uygulanarak toplam kayıp hibrit olarak optimize edilmiştir. Eğitim sırasında öğrenme oranı dinamik olarak azaltılmış ve erken durdurma yöntemi kullanılmıştır. Modelin test verisi üzerindeki doğruluğu %87.32, ağırlıklı ortalama F1 skoru 0.8762 olarak hesaplanmıştır. Sınıf bazlı başarılar incelendiğinde; Derece-0 ve Derece-4 gibi uç sınıflarda yüksek performans (%96.38 ve %94.33 F1) elde edilirken, Derece-2 ve Derece-3 arasında bazı karışıklıklar gözlenmiştir. Bu durum, patologlar açısından da ayrımın güç olduğu sınıflar arasında, modelin benzer zorluklar yaşadığını göstermektedir. ROC eğrisi analizlerinde tüm sınıflar için AUC değerleri 0.97–0.99 aralığında bulunmuş, bu da modelin sınıflar arasında güçlü bir ayırt etme yetisine sahip olduğunu ortaya koymuştur. Ek olarak, sınıflar arası sıralı yapıyı dikkate alan MAE (0.1479) hesaplanarak modelin sınıflandırma başarısı değerlendirilmiştir. Sonuçlar, önerilen modelin hem doğruluk hem de klinik anlamlılık açısından tatmin edici bir performans sergilediğini ve bilgisayar destekli tanı sistemleri içerisinde yer alabilecek potansiyele sahip olduğunu göstermektedir. Sonuçlar, bu çalışmanın özellikle Derece-0 ve Derece-4 gibi görsel olarak belirgin sınıflarda yüksek başarı sağladığını; Derece 1–2–3 gibi birbirine yakın sınıflarda ise insan gözlemlerine benzer zorluklar yaşadığını göstermektedir. Bu durum, modelin klinik karar destek sistemlerine entegre edilebilir potansiyele sahip olduğunu göstermektedir.
Özet (Çeviri)
Renal cell carcinoma (RCC) is the most common histological subtype of kidney cancer in adults, accounting for approximately 85% of malignant renal tumors. According to the Global Cancer Statistics (GLOBOCAN) 2022 report, kidney cancer ranks 14th in incidence and 16th in mortality among all cancers worldwide, with over 430,000 new cases and approximately 156,000 deaths annually. RCC typically originates from the epithelial cells of the proximal renal tubules and is often diagnosed incidentally through imaging modalities such as ultrasound, CT, or MRI performed for unrelated reasons. This silent progression makes early detection challenging, and thus histopathological evaluation of biopsied or surgically resected tissue remains the gold standard for diagnosis. RCC comprises several histological subtypes but the most prevalent being is clear cell RCC (ccRCC), followed by papillary RCC (pRCC) and chromophobe RCC. While histological subtyping has prognostic value, tumor grading—independent of subtype—is a crucial factor in predicting tumor behavior and guiding treatment strategies. Tumor grading reflects how much the cancer cells differ from normal renal epithelial cells and correlates with tumor aggressiveness. Traditionally, RCC grading was performed using the Fuhrman grading system, which evaluates nuclear size, shape, and nucleolar prominence. However, due to reproducibility concerns and limited prognostic clarity, this system has been largely replaced by the WHO/ISUP grading system. Adopted in 2016, the WHO/ISUP system relies primarily on nucleolar prominence as observed at different magnifications (400x and 100x), offering a more standardized and clinically relevant framework. Grades range from 1 (nucleoli absent or inconspicuous) to 4 (extensive nuclear pleomorphism or extreme nucleolar prominence), with Grade-0 referring to normal tissue. This histological grading is essential because it directly influences the prognosis and informs treatment planning, such as the choice between surveillance, surgery, or adjuvant therapy. High-grade tumors tend to behave more aggressively, with increased potential for metastasis and recurrence. Therefore, accurate and reproducible grading is extremely important in clinical decision making. In this study, we propose a novel multi-task deep learning model designed to automatically classify RCC grades from histopathological images. The model architecture is built on the encoder portion of the U-Net framework and enhanced with attention mechanisms that guide the model to focus on diagnostically important regions of interest, particularly nuclear regions in histology slides. The network performs segmentation (of nuclei) and classification (of tumor grade) simultaneously, utilizing a joint learning strategy to improve generalization. We hypothesize that integrating segmentation masks that emphasize nuclear morphology can guide the attention of the model and improve classification performance. To train and evaluate our model, we utilized the publicly available KMC RCC histopathology dataset. The dataset contains H&E-stained histopathological images collected from renal biopsies at Kasturba Medical College (KMC), Manipal, India. The dataset includes 4077 image patches of size 224×224, categorized into five classes: Grade 0 (normal), Grade 1, Grade 2, Grade 3, and Grade 4, based on the WHO/ISUP grading system. The dataset was split into training, validation, and test subsets in an 80:10:10 ratio, ensuring class balance in all sets. During the preprocessing stage, channel-wise normalization was applied to all images using the mean and standard deviation computed from the training set. Additionally, grayscale binary masks representing nuclei regions were generated using color deconvolution techniques to isolate the hematoxylin channel. These masks were used both in the segmentation task and to modulate the attention mechanism during training, allowing the model to learn to prioritize nuclear regions when making classification decisions. The proposed model consists of several core components: an encoder path inspired by U-Net's contracting path, squeeze-and-excitation (SE) blocks for channel-wise attention, a decoder path that produces binary segmentation masks, and fully connected layers for final grade classification. The model receives dual inputs—an RGB image and its corresponding nucleus mask—and outputs both a predicted grade and a binary segmentation map. The loss function used in training is a combination of cross-entropy loss for the classification task and binary cross-entropy (BCE) loss for the segmentation task, balanced by a weighting factor λ=0.7. We used the AdamW optimizer with an initial learning rate of 10^(-4) and applied ReduceLROnPlateau to dynamically decrease the learning rate when validation loss plateaued. Early stopping was implemented to prevent overfitting. The model was trained for 65 epochs, and its performance was evaluated on the unseen test set. The results are promising: the model achieved an overall test accuracy of 87.32%, with a macro average F1-score of 0.8762. Class-wise performance showed particularly high precision and recall for Grade-0 and Grade-4 (F1-scores of 0.96 and 0.94, respectively), which are visually more distinct in terms of nuclear morphology. In contrast, the model showed relatively lower performance in distinguishing Grades 2 and 3, which are morphologically closer and present difficulties even for expert pathologists. Specifically, F1-scores for Grade-2 and Grade-3 were 0.75. To better understand the discriminative power of the model, we computed the Receiver Operating Characteristic (ROC) curves and calculated the Area Under the Curve (AUC) values for each class. All classes achieved AUCs ranging from 0.97 to 0.99, confirming the model's ability to distinguish between different grades with high confidence. The confusion matrix analysis further revealed that most misclassifications occurred between adjacent grades, particularly between Grades 2 and 3 and between Grades 3 and 4, which aligns with known diagnostic uncertainties in clinical practice. In addition to conventional classification metrics, we computed two evaluation metrics specifically suitable for ordinal classification tasks: Mean Absolute Error (MAE). These metrics take into account the ordinal nature of RCC grading and penalize errors based on their severity. The model achieved an MAE of 0.147, indicating good alignment between predictions and ground-truth grade labels. These results validate the model's capability to capture the ordered structure of RCC grades beyond categorical classification. To gain further insights into the decision-making process of the model, we applied Grad-CAM (Gradient-weighted Class Activation Mapping) visualizations to a subset of test samples. These heatmaps showed that the model consistently focused on nuclear regions when predicting higher grades, demonstrating that it had learned to associate morphological features with RCC severity in a clinically meaningful manner. This explainability is particularly important in the context of medical AI, where transparency and trust are critical for adoption in clinical settings. In summary, the proposed Mask-Supervised Attention Classifier demonstrates strong potential as a clinical decision support tool for automated RCC grading. The model achieves high accuracy and robustness, especially in distinguishing between grades with distinct morphological features. While there are limitations in separating morphologically similar grades, these are consistent with human diagnostic challenges. The inclusion of segmentation guidance and attention mechanisms allows the model to focus on histologically relevant regions, contributing to both interpretability and performance. Future work may focus on extending the model to incorporate multi-modal data such as immunohistochemical staining or genetic markers, as well as validating its performance on external datasets from different clinical sources. Additionally, integration into real-time pathology workflows and prospective validation with expert pathologists would further establish its clinical utility. The model's lightweight architecture and high performance make it a promising candidate for practical implementation in digital pathology systems.
Benzer Tezler
- Renal hücreli karsinomlarda CD47 ekspresyonunun prognostik parametreler ile ilişkisinin değerlendirilmesi
Evaluation of the relationship between CD47 expression and prognostic parameters in renal cell carcinomas
ÖMER FARUK DİZİBÜYÜK
Tıpta Uzmanlık
Türkçe
2019
PatolojiGaziantep ÜniversitesiPatoloji Ana Bilim Dalı
PROF. DR. METİN KARAKÖK
DR. ÖĞR. ÜYESİ ZEHRA BOZDAĞ
- Fuhrman nükleer derecelendirmesine göre düşük ve yüksek dereceli böbrek hücreli karsinom ayrımında 3 tesla manyetik rezonans difüzyon ağırlıklı görüntüleme ve dinamik kontrastlı mr incelemenin yeri
Başlık çevirisi yok
TUĞÇE AĞIRLAR TRABZONLU
Tıpta Uzmanlık
Türkçe
2017
Radyoloji ve Nükleer TıpKocaeli ÜniversitesiRadyoloji Ana Bilim Dalı
PROF. DR. ERCÜMENT ÇİFTÇİ
- Estimating metabolic flux variability with machine learning
Metabolik akış değerlerinin makine öğrenmesi ile tahmini
BARIŞ CAN
Yüksek Lisans
İngilizce
2025
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
DOÇ. DR. ALİ ÇAKMAK
- Renal hücreli karsinomda RING box-1 proteinin (RBX1) ekspresyonunun prognostik değeri
Prognostic value of expression ring box-1 proteinin (RBX1) in renal cell carcinoma
EMRE ALTINTAŞ
- Pediatrik abdominal kitlelerde difüzyon ağırlıklı manyetik rezonans görüntülemenin yeri
The role of diffusion weighted magnetic resonance imaging in the diagnosis of pediatric abdominal masses
BETÜL MERVE MEDİKOĞLU
Tıpta Uzmanlık
Türkçe
2019
Radyoloji ve Nükleer TıpErciyes ÜniversitesiRadyodiagnostik Ana Bilim Dalı
PROF. DR. ABDULHAKİM COŞKUN