Authors:

Rachel M Rivera

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

As imaging equipment advances, the resolution of 3D images - such as those of zebrafish embryos - continues to increase, improving detail and accuracy but significantly inflating file size and processing time. NISNet3D, a neural network model optimized for performing image segmentation on 3D images of zebrafish embryos, struggles to handle these high-resolution images efficiently. This research develops a preprocessing and postprocessing pipeline that enables NISNet3D to handle large files more effectively. The preprocessing stage includes Gaussian filtering to suppress noise and Contrast Limited Adaptive Histogram Equalization (CLAHE) to reveal faint cellular boundaries. Large 3D images were handled by splitting them into overlapping chunks and batching them with folders. These folders were then evenly divided between two GPUs, which process their assigned folders sequentially - starting the next one as soon as the current finishes. This balances load and maximizes output rate without exceeding memory limits. After inference, postprocessing stitches the predictions back into one large image. However, closely packed or touching cells in chunk border sections were often mislabeled as a single object. To address this, cell separation was performed using Euclidean distance transforms and watershed segmentation. This method locates each cell's centroid and separates overlapping cells along the line of minimal distance, preserving natural boundaries. Tests using older zebrafish embryo volumes confirm that preprocessing and chunk-based inference significantly reduce runtime while preserving segmentation quality. Visual inspection shows most cells are correctly labeled, though further automation is needed. Future work involves generalizing the pipeline to accommodate other datasets. Keywords: Zebrafish; Image Processing

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Rachel M Rivera