A Lightweight Framework for Semantic Segmentation of Biomedical Images

Abstract

Abstract We introduce a lightweight framework for semantic segmentation that utilizes structured classifiers as an alternative to deep learning methods. Biomedical data is known for being scarce and difficult to label. However, this framework provides a lightweight, easy-to-apply, and fast-to-train approach that can be adapted to changes in image material though efficient retraining. Moreover, the framework is able to adapt to various input sizes making it robust against changes in resolution and is not tied to specialized hardware, which allows efficient application on standard laptops or desktops without GPUs. We benchmark two distinct models, a single structured classifier and an ensemble of structured classifiers, against a U-Net, evaluating overall performance and training speed. The framework is versatile and can be applied to multi-class semantic segmentation. Our study shows that the proposed framework can effectively compete with established deep learning methods on diverse datasets in terms of performance while reducing training time immensely.