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 C

Poster G5 - Mapping Mitochondria Using Self-Supervised Learning and 3D Segmentation AI Techniques

Friday, October 25, 2024

10:00 AM – 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Sophia Hibner (she/her/hers)

Student
Johns Hopkins University
Baltimore, Maryland, United States

Co-Author(s)

LB

Laura Brattain

Associate Professor
University of Central Florida College of Medicine, United States
KC

Katelyn Canedo

Graduate Student
University of Central Florida, United States
KY

Kyu Young Han

Associate Professor
University of Central Florida Creol, United States

Introduction: To develop effective treatment for neurodegenerative disorders such as Parkinson’s and Alzheimer’s disease, an enhanced understanding of brain cell activities is necessary as it can illuminate certain aspects of brain function that are presently evasive, potentially revealing new therapeutic targets and mechanisms of disease progression. Mitochondria are essential to cellular metabolism and their dysfunction is often associated with various neurodegenerative diseases. Through studying mitochondria activity we hope to glean insight into brain function and the first step is to be able to track their movement. This research aims to develop advanced AI algorithms to automatically trace mitochondria in neuronal cells by leveraging self-supervised learning and image segmentation AI techniques.

Materials and

Methods: Two channel brain cell images were collected using widefield microscopy. The red fluorescence channel imaged the microtubule structures and the green fluorescence channel imaged the mitochondria. A deconvolution algorithm was then applied to the images to enhance the edges of the structures. In this study we used the deconvolved green channel of the mitochondria for the segmentation task and a total of 12 images were used for proof-of-concept.
We used the Fiji Trainable Weka 3D Segmentation algorithm to segment the mitochondria. One image was used to train the segmentation model, while the other 11 images were used for inference. The images were processed prior to segmentation by applying Contrast Limited Adaptive Histogram Equalization (CLAHE).
To train a Weka classifier, mitochondria were manually annotated inside the Weka user interface as the foreground while non mitochondria regions were manually annotated as the background (Fig 1). Random forest algorithms were used to select the features from the annotations and train the classifier. Out-of-bag (OOB) error is used to evaluate the model performance.

Results, Conclusions, and Discussions: Results
Fig 2 Left is an example mitochondria image while Fig 2 Right is the Weka segmentation results. Visually we can see that the result matches the original morphology. Fig 3 shows the probability map. The model achieved an OOB error rate of 0.105%. Additionally, qualitative visual evaluation confirmed that the morphology was consistent with the expected structure.
Conclusions
We applied the Trainable Weka Segmentation model to microscopy data to automatically segment mitochondria. The model was trained using one image and then applied it to the remainder of the data set. The results demonstrate the potential of AI-driven methodology to accurately trace mitochondrial structures within neuronal cells. The low OOB error rate indicates the potential of the model, while qualitative assessments reinforce the accuracy of the segmentations produced. This approach not only expedites the tedious process of annotating biomedical images but also has the potential to precede more detailed studies of mitochondrial behaviors in neurodegenerative diseases. By improving our understanding of morphological dynamics in neurons, this research may contribute to the identification of novel biomarkers and therapeutic strategies, ultimately aiding in the development of treatments for neurodegenerative disorders.
Discussions
As shown by the OOB error rate and visual assessment, the model worked reasonably well across various images. Additionally the model offers a more efficient alternative to manual annotation which is often labor intensive and tedious. One weakness of the model is that it tends to lump in multiple individual mitochondria into one segment. The deficiencies of this model leave room for improvement in the future, which may include the integration of more advanced AI segmentation algorithms to differentiate small regions more precisely, as well as using attention mechanisms for better segmentation performance. Additionally there is potential for exploring Generative Adversarial Network (GAN) for synthetic data generation to increase the size of the annotated dataset.

Acknowledgements (Optional):