Detection tracking and classification of large intestine polyps using deep learning methods

Deep learning-based algorithms update millions of parameters during the training phase. The excess amount of parameters requires a large amount of labeled data for training. This restricts the use of high-performance deep learning algorithms in vital areas where labeled data are scarce, such as medical practice. Also, in some applications, even if a sufficient amount of labeled data is available, the data will be divided into classes and reduced. This thesis focuses on improving object detection performance in order to improve object tracking performance in medical images. For this purpose, a new strategy has been developed to reduce the FP rate in object detection. In this strategy, the objects detected by deep learning object detection algorithms are separated from the image and classified. In the classification stage, a method which is independent of the number of images problem, does not require training and which improves tracking performance was proposed. In the proposed method, the feature vector is obtained by passing the broken image through pre-trained CNN architectures. Feature vectors derived from different CNN architectures of the image are combined to reduce the feature extraction time of the CNN architectures without degrading the classification performance. This feature vector was subsampled with DWT and reduced and amplified and classified with DVM. In this context, a new classification method has been introduced to the literature. In order to increase the classifier performance by reducing the size of the CNN feature vector, optimization algorithms and basic filter methods were applied to the CNN features and their performances were compared with DWT. DWT has been shown to contribute significantly to achieving highly promising results in reducing the feature vector size. In addition, the obtained feature vector is used in data matching to improve tracking performance. The proposed method was tested with detection, tracking and classification of large intestinal polyps in conoscopy videos obtained from the public data set, ColonoscopicDataset. According to our researches, the software which classifies three different polyp types has been realized for the first time with this thesis.