Presentation Type
Poster
Presentation Type
Submission
Keywords
Ecology, Ferns, Xylem, Machine Learning, Deep Learning, Semantic Segmentation
Department
Biology
Major
Computer Science/Mathematics
Abstract
Studies of plant anatomical traits are essential for understanding plant physiological adaptations to stressful environments. For example, shrubs in the chaparral ecosystem of southern California have adapted various xylem anatomical traits that help them survive drought and freezing. Previous studies have shown that xylem conduits with a narrow diameter allows certain chaparral shrub species to survive temperatures as low as -12 C. Other studies have shown that increased cell wall thickness of fibers surrounding xylem vessels improves resistance to water stress-induced embolism formation. Historically, these studies on xylem anatomical traits have relied on hand measurements of cells in light micrographs, but this approach is time- and labor-intensive. Here we propose that deep learning-based models can be used to rapidly detect, classify, and measure plant cells with high precision and accuracy. Our goal was to develop models that can detect and classify plant cell types with greater than 95% accuracy.
In this project, we constructed a deep convolutional neural network (DCNN) to segment and classify cell types in light micrographs. We created an encoder-decoder U-Net architecture, where we used convolutional layers to encode the features of the cross section, and transposed convolutional layers to upscale the features to a vessel segmentation mask. We interleaved batch normalization and max pooling layers inside the encoder-decoder blocks to provide a strong regularization to the U-Net. For classification, we explored various transformers and convolutional neural networks to achieve a cell type classification accuracy of 98.1%.
The testing samples were isolated from the training data, and our DCNN performed vessel segmentation on this dataset with high pixel classification accuracy (97.05%) and excellent precision score (80.71%) that represents the model’s ability to predict positive vessel-class pixel values. With further development, the DCNN may provide the ability to measure vessel thickness and area, while also potentially measuring vessel cell wall thickness by performing a digital subtraction of a cell wall mask and vessel mask. This approach could provide opportunities to rapidly analyze larger plant anatomy datasets, allowing us to scale up questions relating plant xylem structure and function to the level of ecosystems or the globe.
Faculty Mentor
Helen Irene Holmlund, Fabien Scalzo
Funding Source or Research Program
Keck Scholars Program
Start Date
24-3-2023 2:00 PM
End Date
24-3-2023 4:00 PM
Included in
Deep learning can be used to classify and segment plant cell types in xylem tissue
Studies of plant anatomical traits are essential for understanding plant physiological adaptations to stressful environments. For example, shrubs in the chaparral ecosystem of southern California have adapted various xylem anatomical traits that help them survive drought and freezing. Previous studies have shown that xylem conduits with a narrow diameter allows certain chaparral shrub species to survive temperatures as low as -12 C. Other studies have shown that increased cell wall thickness of fibers surrounding xylem vessels improves resistance to water stress-induced embolism formation. Historically, these studies on xylem anatomical traits have relied on hand measurements of cells in light micrographs, but this approach is time- and labor-intensive. Here we propose that deep learning-based models can be used to rapidly detect, classify, and measure plant cells with high precision and accuracy. Our goal was to develop models that can detect and classify plant cell types with greater than 95% accuracy.
In this project, we constructed a deep convolutional neural network (DCNN) to segment and classify cell types in light micrographs. We created an encoder-decoder U-Net architecture, where we used convolutional layers to encode the features of the cross section, and transposed convolutional layers to upscale the features to a vessel segmentation mask. We interleaved batch normalization and max pooling layers inside the encoder-decoder blocks to provide a strong regularization to the U-Net. For classification, we explored various transformers and convolutional neural networks to achieve a cell type classification accuracy of 98.1%.
The testing samples were isolated from the training data, and our DCNN performed vessel segmentation on this dataset with high pixel classification accuracy (97.05%) and excellent precision score (80.71%) that represents the model’s ability to predict positive vessel-class pixel values. With further development, the DCNN may provide the ability to measure vessel thickness and area, while also potentially measuring vessel cell wall thickness by performing a digital subtraction of a cell wall mask and vessel mask. This approach could provide opportunities to rapidly analyze larger plant anatomy datasets, allowing us to scale up questions relating plant xylem structure and function to the level of ecosystems or the globe.