| Abstract [eng] |
A biological membrane is a separating layer which forms the outer boundary of a living cell. It is responsible for vital functions such as protecting the cell, regulating the transport of substances in and out of the cell, receiving chemical messengers from other cells and acting as a receptor. In nature, biological membranes can be damaged by toxins which cause the formation of defects on the membrane and due to this reason toxins are often the pathogens of dangerous diseases. Automated defect detection can help diagnose formidable diseases caused by toxins faster. One of the ways to analyse biological membranes is images from atomic force microscopy. Automated defect detection problem could be solved by applying object detection methods on atomic force microscopy images. This paper is composed of six chapters. In chapter 1 a review of related scientific literature is provided. Chapter 2 consists of theory related to the biological membrane, defects of biological membranes and atomic force microscopy. Chapter 3 has the theory of image preprocessing. Chapter 4 discusses the theory related to the methods of defect detection which are applied in this research. Chapter 5 provides a theory of metrics for the object detection task. Chapter 6 analyses the practical part of the research and discusses the results. At the end of the paper plans for further work are provided. The task of detection of defects in atomic microscopy images was carried out by using non-artificial intelligence methods (Hough circle transformation algorithm, Suzuki contour finding algorithm) and convolutional neural networks of various architectures (YOLOv4, CenterNet HourGlass104 512x512, EfficientDet D0 512x512, SSD ResNet152 V1 FPN 640x640, Faster R-CNN Inception ResNet V2 640x640). Data was prepared with augmentation and resizing for convolutional neural networks. All aforementioned methods were evaluated with the PASCAL VOC 2012 mAP and MS COCO 2014 metrics. YOLOv4 reached the highest value in PASCAL VOC 2012 mAP with 72,88% and CenterNet HourGlass104 512x512 reached the highest value in MS COCO mAP with 27,2%. When solving the problem of defect detection in atomic force microscopy images of biological membranes, the result of the PASCAL VOC 2012 mAP metric is more important than the MS COCO 2014 mAP. Due to that, the usage of the YOLOv4 model is recommended. Attempts to improve the metrics obtained by the YOLOv4 convolutional neural network by adding a digital image preprocessing method (Canny edge detection, Otsu and Sobel edge detection operator methods were tested) were unsuccessful. |
