Image analytics play a fundamental role in many fields of science, due to the widespread availability of high-resolution image data. However, image texture is a complex and often neglected property if images. The wealth of information contained in image texture can potentially provide hidden insight into complex phenomena studied through image data. In particular, texture directionality contains information about the dominating orientation of entities of interest, such as proteins and organelles whose properties can be closely related to their directionality. We are exploring image analysis and AI for texture directionality detection. The implemented computational tools enable the automated characterization of texture directionality at different resolutions, with application in the biomedical field and beyond.
Texture is a complex image feature. There is no consensus on its formal definition, but it can be generally regarded as the spatial variation of the brightness intensity of the pixels. Basically, texture consists of spatial patterns whose size and directionality can be, in most cases, perceived by humans. Image analytics experts sometimes refer to the smallest detectable entity of which texture patterns consist of as texton.
Many applications rely on the characterization of the directionality distribution of certain entities of interest. For instance, the directionality of protein fibers such as actin and myosin are associated to cell motility, shape changes and other important cell functions. Similarly, the presence and orientation of defects in certain materials provides valuable insight into their manufacturing process. Clearly, automated texture analysis techniques are of key importance for the analysis of the large amount of available image data.
This project focuses specifically on the characterization of the directional component of texture. The main goals are listed as follows:
Build accurate models to characterize image texture directionality.
Implement texture directionality detection tools that enable the quantitative characterization of texture directionality across the image at different levels of resolution.
Explore AI-based approaches for texture directionality detection.
Developed a novel texture directionality detection technique. This technique consists of an interpolation-based model for the computation of the Gray Level Co-occurrence Matrix (GLCM), which enables the characterization of image pixel distribution at different offsets and along different directions. The interpolation-based model enables GLCM computations for any real valued angle and offset, as opposed to the traditional lattice-based model. This allows texture directionality detection at different levels of resolution.
Built a library of synthetic texture images with different blur and additive noise levels. Based on these levels, the robustness of the directionality detection technique was quantified.
The performance of the directionality detection technique was initially demonstrated on fluorescence microscopy images of fibroblast cells containing protein fibers, whose directionality is of interest in many biological studies.
Deep and shallow neural network (DNN and SNN) for texture directionality detection were designed. Synthetic texture images from the above-mentioned library were used to train and test DNN and SNN performance. Current data shows accuracy comparable to GLCM-based directionality detection with higher computational efficiency.
Deep learning (DL) networks were built for the interactive visualization of live-cell images and for the characterization of textural features of interest. The DL-based approach scales to gigapixel microscopy images.