Yapay zeka ile gliomlarda KI67 işaretleme

Determining the stage of cancer is one of the essential steps for treatments. The cellular proliferation index is one of the factors determining the stage of cancer. Therefore, determining the cellular proliferation index is a vital part of treatment. Pathologists use KI67 nuclear proteins' prognostic feature to determine cellular proliferation. High proliferation rate cells have a brownish color when KI67 nuclear protein is applied. Therefore, the quantity of these cells is a crucial part of calculating the proliferation index. In the scope of this thesis, image processing with artificial intelligence techniques is used for labeling cells as KI67 positive and negative. In this study, the Faster R-CNN algorithm was selected for object detection. Faster R-CNN object detectors' accuracy depends on CNN-based feature extractor model. Today, many different CNN models are available for this purpose. In this study, ResNet50, ResNet18, VGG16, VGG19 CNN models were used for feature extractors, and their KI67 cell-labeling performance was tested. Object detection with deep learning techniques depends on the dataset. Dataset images should be labeled before the training process. In this study, two dataset image labeling techniques were implemented: Supervised and Semi-Supervised learning methods. For the Supervised learning method, all dataset images were labeled manually. For the Semi-Supervised learning method dataset, images were labeled with a previously trained Faster R-CNN object detector. Integrating the Semi-Supervised learning system into the Faster R-CNN object detection model, the object-labeling process has achieved more accurate results. In that way, nuclear KI67 expressing glioma cells and non-expressing cells were identified most appropriately.