Makine öğrenimi ile kanserli beyin hücrelerinin tespiti

In recent years, there has been an increase in the incidence of brain tumors. Brain tumors are a serious condition that involves the slow growth of a foreign mass in the brain, which can lead to death. Early and accurate detection of tumors is crucial in combating the disease. MRI imaging technology is used to obtain better patterns of brain tumors compared to other technologies. However, detecting these images is not easy, and in some cases, when the disease is difficult to detect, it may be too late for the patient. At this stage, image processing techniques become very important. In this thesis, image processing techniques are crucial. The data from MRI images are analyzed using segmentation and classification techniques, which provide significant benefits to doctors and the medical world in terms of the segmentation, location, and characteristics of the disease. Thus, it may be possible to diagnose diseases at an early stage. In this thesis, 7023 MRI images labeled as glioma, meningioma, pituitary, and notumor from Kaggle were used. These images were cropped from the top and bottom using specific algorithms and resized to dimensions using various preprocessing methods. Once the MRI images were scaled to the desired dimensions, they were grouped into axial, sagittal, and coronal views to improve accuracy and training performance. This thesis deeply examines the application of an advanced artificial intelligence model called EfficientNet in the classification of brain tumors. EfficientNet, when trained on a large dataset, is an architecture that can distinguish different types of brain tumors with high accuracy. During the training process, the ability of deep learning algorithms to recognize complex structural and visual patterns is crucial for the accurate classification of tumors in MRI images. Examining vital characteristics such as the type, size, and location of tumors plays a central role in this process. In this thesis, the integration of the Convolutional Block Attention Module (CBAM) helps optimize the learning process of the models. CBAM allows the model to select important features and reduce background noise,especially through spatial and channel-based attention mechanisms, significantly enhancing the model's object localization and classification performance. This approach strengthens the model's capacity to analyze brain tumor MRI images more effectively, enabling more precise diagnoses. The focus of the thesis is on distinguishing various subtypes of brain tumors. Categories such as glioma, meningioma, pituitary tumors, and non-tumorous areas are thoroughly examined during the model's training. For each of these categories, special training datasets are used to test the model's accuracy and reliability. Various image augmentation techniques are employed during the preprocessing stage of the datasets using 'ImageDataGenerator'. These techniques include rotation, shifting, scaling, and other image manipulations, helping the model better adapt to real-world variations. Additionally, these techniques reduce the risk of overfitting and increase the overall accuracy of the model. During the training of the model, the KFold cross-validation method is applied. This method allows the model to be trained and tested multiple times using different subsets of the dataset. Each fold evaluates the model's performance on different data subsets, providing more reliable results, especially in cases where data is limited. This technique allows for a more accurate measurement of the model's performance. The model's training process consists of two stages: In the first stage, the layers of EfficientNetB7 are frozen, and only the top layers are trained. This approach allows the model to initially learn general features and establish a stable foundation. In the second stage, some lower layers are also included in the training, and fine-tuning is performed. This fine-tuning process helps the model recognize more specific and complex features, achieving higher accuracy.This thesis demonstrates how machine learning and artificial intelligence can bring significant innovations to the field of medical diagnosis. These technologies, surpassing the limitations of traditional methods, open new horizons for medical research and applications by enabling early diagnosis and rapid treatment. The potential of this innovative approach, particularly in the diagnosis of brain tumors, is comprehensively addressed in this study.