MLOPS APPROACH FOR AUTOMATIC IMAGE SEGMENTATION

Abstract

The development of artificial intelligence and algorithms for classification and image segmentation has significantly contributed to the development of technologies in the field of automatic diagnosis with minimal human participation. The key feature of this type of task is the need to use a large amount of data and the need for significant computing resources. There is also a need to use cloud computing for collaborative work on projects. When analyzing such specific and complex images as immunohistochemical, histological, and cytological, only the use of approach algorithms is insufficient. Therefore, the use of convolutional neural networks with the U-net architecture for automatic segmentation has become widespread. The choice of tools for implementing the pipeline plays an important role in ensuring a clear and efficient team process. Key tools include containerization systems, infrastructure as code delivery tools, and CI/CD platforms such as GitHub Actions. When designing a pipeline for machine learning applications, you should consider the available resources (hardware or cloud). The selection of frameworks for working with data and implementing U-nets should be considered during the automatic segmentation stage. The advantage, and at the same time the limitation, of this system is that it specializes only in automatic segmentation using Unet. However, the image processing stage and the generation of a dataset for segmentation tasks is a distinct feature that distinguishes this system from others.  The main goal of this article is to develop a mechanism for creating, training, and managing neural network models designed for automatic segmentation of immunohistochemical images.

This paper proposes a life cycle structure for biomedical image segmentation based on MLOps practices. A feature of this approach is the development of a mechanism for image processing using image masks.