Biyomedikal imgelerde çapasiz dedektörlerle görsel nesne tespiti

The role of data analysis in health informatics has grown rapidly in the past years due to the increased number of biomedical data. In the context of biomedical images, object detectors play an important role in automating image analysis tasks and facilitating high-throughput analysis of large datasets. Biomedical images can include various types of data, such as histology slides, microscopy images, and medical scans. Object detectors can be trained by using convolutional neural network architectures to detect specific objects within these images, such as cells, nuclei, or tumors. In traditional visual object detectors, anchor boxes are pre-defined boxes of different scales and aspect ratios that are placed on a grid over the image. The detector then predicts the offset of the object's bounding box relative to the anchor box, as well as the object's class. Despite the good performance of state-of-the-art anchor-based models, they require predetermined anchor parameters such as size, number, and aspect ratio of anchors and have limitations when dealing with small nodules having different sizes. To overcome this, we have shifted our research towards the application of anchor-free models in biomedical images. We have selected YOLOX, an anchor-free object detector that automatically predicts the position, center, height, and width of the detected nodule without the design of the anchor parameters. In this thesis, we developed an improved model that includes several modifications to the standard YOLOX model. The Distance IoU loss function is utilized to train the nodule detection efficiently, resulting in faster convergence of the detector. Attention modules were also incorporated into our improved model to select the essential areas of the image and increase detection accuracy. Additionally, we introduced the varifocal loss to solve the class imbalance problem associated with one-stage detectors. Our modified method outperforms both anchor-based and anchor-free models when tested with brain tumor, lung cancer, and kidney stone datasets. After conducting experiments, we achieved a mean average precision of 87.56% for brain tumor detection, 84.11% for the Lung Image Database Consortium dataset, and 83.06% for kidney stone detection.