The influence of CNN architecture, image size and quality to object detection model on histological specimens

Improving convolutional neural network (CNN) quality for object search in histology scans is a long-standing problem that essentially involves selecting the best CNN architecture and creating a high-quality dataset. The efficacy of object detection algorithms is determined by numerous factors, including image quality, image size, and the search object. The primary aim of this study was to identify published studies on the impact of various image characteristics in a training sample and CNN architecture on the quality of a created model. Literature published in the last 5 years was reviewed, which addressed data pre-processing, methodology, requirements to images included in datasets, image preparation for CNN model development, and architecture selection. At the time of the study, there were no requirements to image size, and there was no data on the ratio of object size to image size for the best model performance. Moreover, the selection of neural network architecture is lacking in transparency and algorithmization. In the majority of cases, researchers recommend architectures that they have developed or used themselves, without explaining the reasons and selection criteria or comparing them to alternative options. All these factors significantly complicate the development of CNN models for medical image processing. This paper presents a brief overview of publications that address image preparation for datasets, as well as a potential approach to CNN architecture selection.