Brain Tumor Segmantation With Deep Learning

Bilal Taşdemir; Necaattin Barışçı

Makale detayları

Brain Tumor Segmantation With Deep Learning

Yazarlar:

Bilal Taşdemir | Necaattin Barışçı

Yayıncı:

Gazi University

Ana Başlıklar:

Derin Öğrenme | Yapay Zeka (Diğer)

Dergi:

Bilişim Teknolojileri Dergisi Vol.17 Issue 3, pp.159-174

Belge Tipi:

Makale

Yayın Tarihi:

7/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001323522

pISSN:

1307-9697

Anahtar kelimeler:

Özet:

 With the increasing population, more and more people are affected by brain tumors every day. Compared to other diseases, the death rate of brain tumors is much higher. In addition, people suffering from brain tumor disease have important difficulties in their daily lives. The diagnosis of brain tumors poses a challenging process for medical professionals. To address the difficulties faced by doctors during the diagnostic phase and propose a solution, the objective is to design a model that enables the automatic segmentation of brain MR images. In this study, Ensemble learning method was used to overcome the difficulty in segmentation of brain MRI images. Deep learning base attention u-net and u-net models were used in ensemble learning method. The aim is to develop a more stable model by averaging the prediction values from two different models. While the 2018, 2019 and 2020 versions of BRATS dataset were used for model training, the 2017 version was preferred for model testing. With different data preprocessing steps and ensemble learning model, the difficulty of segmentation of brain MR images has been overcome on imbalanced dataset. With the ensemble learning method created, 87.33 average zar score, 81.74 necrotic class zar score, 91.57 edema zar score, 76.03 enhancing zar score, 99.96 background class zar score and zar score of 83.11, 78.88 and 76.03 for Whole Tumor WT, Tumor Core TC and Enhancing Tumor ET were obtained on BRATS2017 dataset.

 With the increasing population, more and more people are affected by brain tumors every day. Compared to other diseases, the death rate of brain tumors is much higher. In addition, people suffering from brain tumor disease have important difficulties in their daily lives. The diagnosis of brain tumors poses a challenging process for medical professionals. To address the difficulties faced by doctors during the diagnostic phase and propose a solution, the objective is to design a model that enables the automatic segmentation of brain MR images. In this study, Ensemble learning method was used to overcome the difficulty in segmentation of brain MRI images. Deep learning base attention u-net and u-net models were used in ensemble learning method. The aim is to develop a more stable model by averaging the prediction values from two different models. While the 2018, 2019 and 2020 versions of BRATS dataset were used for model training, the 2017 version was preferred for model testing. With different data preprocessing steps and ensemble learning model, the difficulty of segmentation of brain MR images has been overcome on imbalanced dataset. With the ensemble learning method created, 87.33 average zar score, 81.74 necrotic class zar score, 91.57 edema zar score, 76.03 enhancing zar score, 99.96 background class zar score and zar score of 83.11, 78.88 and 76.03 for Whole Tumor WT, Tumor Core TC and Enhancing Tumor ET were obtained on BRATS2017 dataset.

 With the increasing population, more and more people are affected by brain tumors every day. Compared to other diseases, the death rate of brain tumors is much higher. In addition, people suffering from brain tumor disease have important difficulties in their daily lives. The diagnosis of brain tumors poses a challenging process for medical professionals. To address the difficulties faced by doctors during the diagnostic phase and propose a solution, the objective is to design a model that enables the automatic segmentation of brain MR images. In this study, Ensemble learning method was used to overcome the difficulty in segmentation of brain MRI images. Deep learning base attention u-net and u-net models were used in ensemble learning method. The aim is to develop a more stable model by averaging the prediction values from two different models. While the 2018, 2019 and 2020 versions of BRATS dataset were used for model training, the 2017 version was preferred for model testing. With different data preprocessing steps and ensemble learning model, the difficulty of segmentation of brain MR images has been overcome on imbalanced dataset. With the ensemble learning method created, 87.33 average zar score, 81.74 necrotic class zar score, 91.57 edema zar score, 76.03 enhancing zar score, 99.96 background class zar score and zar score of 83.11, 78.88 and 76.03 for Whole Tumor WT, Tumor Core TC and Enhancing Tumor ET were obtained on BRATS2017 dataset.

 Artan nüfus ile birlikte her geçen gün daha fazla insan beyin tümöründen etkilenmektedir. Diğer hastalıklar ile kıyaslandığında beyin tümörünün ölüm oranı çok daha yüksektir. Ayrıca beyin tümörü hastalığına yakalanan bireyler günlük yaşamlarında önemli zorluklarla karşılaşmaktadır. Beyin tümörünün tanısı doktorlar için zorlu bir süreçtir. Teşhis aşamasında doktorların karşılaştığı zorluklar ve bu sorunların üstesinden gelmek adına bir çözüm önerisi olarak, beyin MR görüntülerinin otomatik segmentasyonunu sağlayacak bir modelin tasarlanması hedeflenmiştir. Bu çalışmada beyin MR görüntülerinin segmentasyonundaki zorluğu aşmak için topluluk öğrenimi yöntemi kullanılmıştır. Topluluk öğrenimi yönteminde derin öğrenme tabanlı dikkat mekanizmalı u-net ve u-net modelleri kullanılmıştır. Bu yöntem ile iki farklı modelden gelen tahmin değerlerinin ortalamasının alınması ve daha kararlı bir modelin geliştirilmesi amaçlanmıştır. Model eğitimi için BRATS veri setinin 2018, 2019 ve 2020 versiyonları kullanılırken, model testleri için 2017 versiyonu tercih edilmiştir. Veri setindeki dengesiz sınıf dağılımı problemine çözüm olarak farklı veri ön işleme adımları kullanılmıştır ve topluluk öğrenimi modeli ile beyin MR görüntülerinin segmentasyon problemi ele alınmıştır. Elde edilen topluluk öğrenimi yöntemi ile BRATS2017 veri seti üzerinde 87,33 ortalama zar skoru, 81,74 nekrotik sınıfı zar skoru, 91,57 ödem sınıfı zar skoru, 76,03 artırılmış tümör sınıfı zar skoru, 99,96 arka plan sınıfı zar skoru ve Tüm Tümör TT, Çekirdek Tümör ÇT ve Artırılmış Tümör AT için sırasıyla 83,11, 78,88 ve 76,03 zar skoru elde edilmiştir.

Makale detayları

Brain Tumor Segmantation With Deep Learning

Yazarlar:

Bilal Taşdemir | Necaattin Barışçı

Yayıncı:

Gazi University

Ana Başlıklar:

Derin Öğrenme | Yapay Zeka (Diğer)

Dergi:

Bilişim Teknolojileri Dergisi Vol.17 Issue 3, pp.159-174

Belge Tipi:

Makale

Yayın Tarihi:

7/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001323522

pISSN:

1307-9697

segmantation

tumor

deep learning

u-net

image processing

normalizationSegmentasyon

Tümör

Derin Öğrenme

U-net

Görüntü İşleme

Normalizasyon

Topluluk Öğrenimi

Sayı

Makale detayları

Yazarlar: Bilal Taşdemir | Necaattin Barışçı

Yayıncı: Gazi University

Ana Başlıklar: Derin Öğrenme | Yapay Zeka (Diğer)

Dergi: Bilişim Teknolojileri Dergisi Vol.17 Issue 3, pp.159-174

Belge Tipi: Makale

Yayın Tarihi: 7/2024

Belge Dili: Türkçe

Document ID (Org No.): 2001323522

pISSN: 1307-9697

Sayı

Yazarlar:

Bilal Taşdemir | Necaattin Barışçı

Yayıncı:

Gazi University

Ana Başlıklar:

Derin Öğrenme | Yapay Zeka (Diğer)

Dergi:

Bilişim Teknolojileri Dergisi Vol.17 Issue 3, pp.159-174

Belge Tipi:

Makale

Yayın Tarihi:

7/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001323522

pISSN:

1307-9697