Efficient Volumetric Image Segmentation
The three-dimensional organization of biological entities has a significant impact on disease development, progression and outcome.
In deep learning, the computational complexity of 3D biomedical image segmentation often leads to memory errors. To tackle this issue, a common practice is to train with downsampled images. However, this can lead to accuracy reduction when applying the model on original resolution images. FNOSeg3D [1] is introduced as a 3D segmentation model designed to be robust to variations in training image resolution. It leverages the Fourier neural operator (FNO) [2], a deep learning framework to learn mappings between functions in partial differential equations. FNO has the ability to enhance image resolution without explicit training on high-resolution data, called zero-shot resolution. This approach appears to be a promising advancement in the field of 3D biomedical image segmentation, particularly in terms of addressing the challenges related to image resolution variations and computing resources.
2) Then, the work consists in optimizing and testing the FNOSeg3D model that has been implemented in Pytorch in our lab and make an analysis of this model in the perspective of zero-shot resolution. The transfer of the model to 3D optical images (confocal images) will be appreciated.
[2] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar, “Fourier neural operator for parametric partial differential equations,” in International Conference on Learning Representations, 2021.
