Cnn derin öğrenme tekniği kullanılarak prostat kanser hastalığının teşhisi = Diagnosis of prostate cancer disease using Cnn deep learning technique

Abstract

Prostat kanseri (PKa) için bir risk sınıflandırma yöntemi olan Gleason derecelendirmesi subjektif, yani girdi görüntüyü inceleyen ve rapor eden patoloğun deneyim ve uzmanlığına yoğun bir biçimde bağlıdır. Onun dışında, gleason derecelendirmesi PKa teşhisi/tedavisi için önemli olsa da zaman alıcı bir iştir. Ayrıca, biyopsi görüntülerini teşhis/derecelendirme süreci gözlemci patologlar arasında önemli farklılıklara sahip olup, bireysel hastalar için teşhis/tedavi verimliliği ve etkinliğini sınırlayabilmektedir. Makine Öğrenmesi (MÖ) ve Derin Öğrenme (DÖ) sistemleri, gleason derecelendirmesinin nesnelliğini ve verimliliğini artırma konusunda umut vaat etmiştir. Ancak, derin öğrenme ağları tek başına kullanıldığında, eğitim verileri dışındaki bir kaynaktan Tüm Slayt Görüntülerinde (TSG) etki alanı kayması ve düşük performans sergiler. Dolayısıyla, bu çalışmada makine öğrenmesi ve derin öğrenme algoritmalarının karma olarak kullanılması amaçlanmıştır. Bu sistemde, girdi biyopsi görüntüsü üzerinde önemli bölgeleri otomatik olarak belirleyebilen ve belirlenen bölgeleri doğru bir şekilde sınıflandırabilen yapay zekâ tabanlı prostat kanseri tespit ve teşhis sistemi geliştirilmiştir. Önemli bölgeleri tespit etmek (yerelleştirme görevi için) YOLO algoritması kullanılmıştır. Algoritma, bu çalışma için oluşturulan prostat kanseri veri kümesi ile yeniden eğitilmiştir. Veri seti, 500 gerçek prostat dokusu biyopsi görüntüsünden oluşmaktadır. Veri seti, veri artırma işleminden önce sırasıyla 450/50 gerçek prostat dokusu görüntüsü olarak eğitim/test kısımlarına ayrılmıştır. Daha sonra 450 etiketli biyopsi görüntüsünden oluşan eğitim seti, veri çoğaltma yöntemi ile ön işleme tabi tutulmuştur. Bu sayede veri setindeki biyopsi görüntü sayısı 450'den 1776'ya yükseltilmiştir. Ardından veri seti ile algoritma eğitilmiş ve otomatik prostat kanseri tespit ve teşhis aracı geliştirilmiştir. Geliştirilen araç iki test seti ile test edilmiştir. İlk test seti, eğitim setine benzeyen 50 görüntü içermektedir. Böylece %97 tespit ve sınıflandırma doğruluğu sağlanmıştır. İkinci test seti ise tamamen farklı 137 gerçek prostat doku biyopsi görüntüsü içerir. Böylece %89 tespit doğruluğu elde edilmiştir. Bu çalışmada, otomatik prostat kanseri tespit ve teşhis aracı geliştirilmiştir. Test sonuçları, nesne algılama algoritmaları gibi yapay zekâ (bilgisayar görüşü) yöntemleri kullanılarak yüksek doğruluklu (yüksek performanslı) prostat kanseri teşhis araçlarının geliştirilebileceğini göstermektedir. Bu sistemlerin, patologlar arasındaki gözlemciler arası değişkenliği azaltabileceği ve tanı aşamasındaki zaman gecikmesini önlemeye yardımcı olabileleceği kanısına varılmıştır.

The most precious thing in the realm of existence is life. And the most valuable thing among duties is to serve life. Therefore, it is certain that the health sector has a very important place in our lives. With the advancement of technology, many studies have been and are being carried out to help doctors and bring the health sector to a more advanced level. The biggest thing that affects human health is getting cancer. Unfortunately, many people all over the world and in our country die due to cancer. According to the statistics, lung cancer is known as the most common type of cancer in the world today. Prostate cancer comes second. The risk of developing cancer is one of the leading health problems in Turkey as well as all over the world. According to the Turkish Cancer Statistics data of the Ministry of Health, prostate cancer ranks second among the top 10 cancer types in 2015. When considered according to all age groups, prostate cancer ranks second with 12.9%, 13.2% when considered according to 50-69 age groups, and 18.3% when considered according to age groups of 70 and above. According to the results of the research, it is obvious that determining the causes of the disease and taking measures against cancer in order to protect people from cancer, which affects human health the most, will contribute greatly to human health and the future of the country. Considering these great goals, it is seen that it has become mandatory for the information technologies and the health services community to work together. This is why interdisciplinary work has an important place in the fight against cancer. Working between disciplines not only saves time but also enables a more accurate diagnosis and treatment to be made. Detecting prostate cancer is a difficult and time-consuming task performed manually by pathologists. Diagnosis of pathologists according to the degree of disease; may vary according to personal experience, volume of practice and natural subjectivities. When diagnosing the disease in prostate cancer, our biggest predictor is to determine the gleason score and tumor stage. The Gleason grading system is considered part of a standard protocol when evaluating prostate histopathological specimens (biopsy). Grading is based on tumor appearance and architectural patterns shown in prostate tissue samples in glands. The rating score ranges from 1 to 5. The development of methods based on digital image processing and machine learning to classify tissues and determine the degree of disease, as well as predict the outcome of disease and automatically analyze pathology images, has become popular recently. Computer-assisted pathology has made it easier for pathologists to diagnose tissue samples. The deep learning method, which is a supervised learning technique by extracting automatic features from the data, where machine learning and classical probabilistic and statistical methods are insufficient, has become quite common in diagnosing many types of cancer. The CNN architecture of deep learning consists of three layers: the input layer, the convolution layers, and the fully connected classification layer. CNN architecture; It consists of widely used object detection and classification sub-architectures such as RCNN, FR-CNN, DenseNet-121, ResNet-50, ImageNet, VGG-16, MobileNet SSD, Inception-V3 and YOLO. There are many studies using CNN architecture algorithms for the diagnosis of prostate cancer disease. The fact that these studies obtained different results from each other is due to the fact that the model is different, the data set has a different set, and the amount of data set of the trained system is low. In the past, the accuracy of the systems in the detection and diagnosis of diseases using CNN techniques was generally below 60%. Therefore, the difference and low accuracy rate of these studies is due to many factors and problems such as the lack of object detection part of some methods, the small data set and the training process. Methods such as MobileNet SSD, DenseNet-121, ResNet-50, FR-CNN, R-CNN do not scan the entire input image, but focus only on randomly selected regions. As a result, these methods may miss important regions. This may result in misclassification. However, the YOLO object detection method detects and classifies objects or regions by scanning an entire input image. For this, the YOLO method has a very powerful structure in detecting objects or desired regions. The first of the reasons why our study differs from other studies is that there is not a new method with the YOLO method and there is no study conducted with YOLO in the diagnosis of prostate cancer disease so far. The second major difference is that these models, together with the previous probabilistic and metaheuristic algorithm methods, diagnose prostate cancer with very few data sets in the diagnosis of disease. In our study, 1776 datasets were used. The subject of this thesis is to find the position of healthy or unhealthy tissues in an input biopsy image and to classify the found region into classes such as benign, grade 3, grade 4, and grade 5 by grading with the Gleason score. We performed these operations by retraining the YOLO-based general-purpose object detection algorithm with our own dataset samples. As a result, an automated prostate biopsy image processing and diagnostic tool was developed. We realized the performance and accuracy values of this developed system by using complex matrix (confusion matrix) and accuracy (accuracy) measurement methods. In our system, the YOLO algorithm was retrained with 1776 prostate tissue pattern images obtained from 450 real patient tissue images. Relevant sites were labeled by one pathologist and confirmed by two different pathologists. The performance of the system was measured with two test sets. Test set 1 consists of 50 biopsy images similar to the images in the training set. Test set 2, on the other hand, consists of completely different biopsy images from the images in the trained system. All images in test set 1 and test set 2 were carefully reviewed by three pathologists. As a result, an average of 97% accuracy was obtained for test set 1 and an average of 89% accuracy for test set 2.The results of the system show that an artificial intelligence system based on the YOLO object detection algorithm can achieve successful detection and classification results between benign biopsy nuclei and nuclei containing malignancy. It can also be concluded that an artificial intelligence system can grade prostate biopsies with high performance. Increasing the number of samples in order to shed light on future studies may yield more positive results as it will increase the reliability and performance of the system. In addition, better learning of the cellular properties of the system by going down to the core details of the tissues with closer photos can increase the performance and accuracy of the system.