The swift progress in X-ray physics and the expanding capabilities of X-ray synchrotron sources have made the analysis of tomographic X-ray datasets increasingly vital in various scientific, medical, and industrial fields.
This project deals with the segmentation of images depicting the microstructure change of solid oxide cells (SOC). The manual segmentation of these images, however, is a labor-intensive and error-prone task, often riddled with uncertainties and subjectivity. Automating the segmentation process is crucial to facilitate scientific findings and industrial insights.
In the current landscape of medical image segmentation, two prominent deep neural network structures stand out: U-net++ [1] and VGG [2]. U-net++, recognized for its effectiveness in capturing intricate spatial features, has become a go-to choice. Meanwhile, VGG, renowned for its deep convolutional architecture, continues demonstrating robust performance in various medical imaging applications. These state-of-the-art DNN structures play a pivotal role in advancing the accuracy and reliability of medical image segmentation.
A possible direction for our project is combining UNet++ [1], VGGNet [2], and some characteristics of the Boundary-Sensitive Network (BSN) [3], which could lead to an advanced and highly effective model for semantic segmentation tasks. To the best of our knowledge, this particular combination has yet to be attempted and is not present in literature. Therefore, it could lead to exciting results. However, if it proves too challenging, we will consider switching to a less complex architecture, e.g., removing the BSN part of the implementation.
[1] Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, e J. Liang, «UNet++: A Nested U-Net Architecture for Medical Image Segmentation». arXiv, 18 luglio 2018. Consultato: 29 ottobre 2023. [Online]. Disponibile su: http://arxiv.org/abs/1807.10165
[2] K. Simonyan e A. Zisserman, «Very Deep Convolutional Networks for Large-Scale Image Recognition». arXiv, 10 aprile 2015. doi: 10.48550/arXiv.1409.1556
[3] X. Du et al., «Boundary-sensitive Network for Portrait Segmentation», in 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), mag. 2019, pp. 1–8. doi: 10.1109/FG.2019.8756516