Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session A

Poster F5 - A Machine Learning Model for Identifying Mitochondria Fusion and Fission Points

Thursday, October 24, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Tallah Modirzadeh

Student
San Jose State University
San Jose, California, United States
RM

Raul S. Michael

Student
San Jose State University
Folsom, California, United States

Co-Author(s)

IT

Issa Tahir

Professor
San Jose State University, United States

Primary Investigator(s)

Patrick Jurney, PhD

Assistant Professor
San Jose State University, United States

Introduction: Understanding the physiological processes underlying cardiovascular disease (CVD) involves examining endothelial cell (EC) mitochondrial networks, as mitochondrial function and structure are crucial for EC metabolism, thereby impacting CVD progression. Currently, biochemical assays and immunofluorescence (IF) microscopy provide insights into how mitochondrial structure and function influence cellular metabolism, but lack the ability for live observation of mitochondrial networks. Therefore, the dynamics of mitochondrial network organization in response to mechanical or chemical stimuli are difficult to study and poorly understood. Holotomographic microscopy (HTM) is a promising technique for real-time, label-free visualization of ECs and specifically, their intracellular organelles such as mitochondria. This non-destructive and non-interfering live cell imaging method offers opportunities to observe mitochondrial network dynamics. However, we show that these existing processing tools, based on IF microscopy techniques, are incompatible with HTM images, calling for the utilization of a machine-learning model capable of utilizing deep learning to accurately identify mitochondrial networks and their dynamics. Our model uses a Unet learner with a Resnet18 encoder to identify four classes: EC cytoplasm, EC border, substrate, and EC mitochondria. As no current tool allows for HTM image segmentation, this model allows for accurate segmentation of HTM EC images outlining the four key components of HTM images of ECs. By enabling the study of mitochondrial networks and their effects on CVD, this approach offers a new tool to expand our understanding of the disease mechanisms.

Materials and

Methods: Images were taken using a CX-F HTM (Nanolive) with Human Umbilical Vein Endothelial Cells (HUVECs) cultured in 35mm (7mm side wall, polymer coverslip bottom) Ibidi dishes. In the acquisition process, cultured cells were placed on the stage top and a raw image z-stack was acquired. Utilizing ImageJ macros, the stack is compressed into a max image projection (MIP). Upon identification of an EC, the image was imported through the microscope software to a remote laptop for further analysis. We then used a tablet to hand annotate the mitochondrial networks and cell borders and used Adobe Photoshop to identify the EC cytoplasm and substrate classes.
Our model was trained using the approach outlined in Figure 1A utilizing 145 hand-annotated HTM images. These images were imported into our model for augmentation and iterative training, where the model continuously updates input parameters based on the errors made during training and validation. To address the class imbalance within the images, a weighted loss function was implemented to prioritize the identification of mitochondria and the cell border. Training and validation loss assessed the model’s ability to learn and generalize new data; paired with an accuracy function, this determined model training performance. Additionally, we implemented an identifier that saved the highest quality model based on the validation loss and accuracy metric. Upon completion of model training, five test images were utilized to assess the accuracy of the trained model as shown in Figure 1B.

Results, Conclusions, and Discussions: We report a ML model that utilizes segmentation of ECs through training in four classes: EC cytoplasm, EC border, substrate, and EC mitochondria. The training results of the model are shown in Figure 2, where the training and validation loss converge at epoch 4 with an accuracy of 0.962. We successfully segment the test image with high accuracy as indicated by the accuracy metric of 94.17% in Figure 3. The structural similarity index measurements (SSIM) of the test results were computed to compare the model’s segmentation quality to other microscopy analysis tools as depicted in Figure 4. This metric determined how similar an image is to its reference image and allowed for quantification of the accuracy of segmentation. In Figure 4, the model output yields an average SSIM value of 0.918704 while the Mitochondrial Analyzer plugin yields an SSIM value of 0.20165. Correspondingly, we use the CellProfiler software with the identical image, calculating an output image SSIM value of 0.47057.
When inputting a test image, the model yields an SSIM value of 0.91891 which is 128.018% higher than the ImageJ Mitochondrial Analyzer plugin SSIM value. Similarly, the CellProfiler software yielded an SSIM value of 0.47057 which is 64.5335% lower than that of the model. Both of these SSIM values indicate the model’s close similarity between test and output images whereas the Mitochondrial Analyzer plugin and CellProfiler displayed low similarity as shown in Figure 4.
We can improve the accuracy of the model as qualitative indications of overfitting appear present in the model output image. This overfitting affects the similarity between the input and the output image and thereby the SSIM value. To address this, increasing the training dataset and optimizing the weighted loss function will allow the model to segment accurately. After optimizing the model for higher accuracy, we will implement the model to quantify fusion and fission events in live endothelial cells under various mechanical and chemical stimuli in order to study their role in CVD.

Acknowledgements (Optional):