Deep UNet for Mitochondira Segmentation (Zeiss India Contest)

Members: Rohith Srinivaas M, Rishhanth Maanav V, Lokesh Kumar T (IIT Madras)

Mitochondria - Power Factory of the Cell

• Morphology and distribution of intercellular components are of substantial biological importance
• Abnormal Mitochondria morphology can be seen in Parkinson’s disease related genes.
• Accurate Mitochondria segmentation can improve cell segmentation accuracy

Deep Learning in Microscopy Imaging

• Deep Learning revolutionized the way Computer Vision has been seen in recent years.
• Deep learning has been well exploited in the field of Medical Imaging.
• We have used state of the art recently developed techniques for Mitochondria segmentation.

Deep Unet for Image Segmentation

• Unet - Deep Learning Architecture used to solve biomedical image segmentation tasks.
• An improved version of Unet - Deep Unet, (Sept. 2018) was originally used for Pixel level sea land segmentation from satellite image.
• We have used Deep Unet for Mitochondria segmentation.

Dataset Description

• The dataset used is Electron Microscopy dataset.
• The dataset represents a 5x5x5µm section taken from the CA1 hippocampus region of the brain, corresponding to a 1065x2048x1536 volume.
• The resolution of each voxel is approximately 5x5x5nm.
• The data is provided as multipage TIF files.

Training Details

• The dataset was divided into train(3000 images) and test(500 images).
• 6-fold cross-validation was performed during the train procedure for better generalisation of the model.
• Libraries used: TensorFlow, Keras
• Training was performed with GTX 1080 Ti for 6 hrs..
• The training iteration was continued for 150 epoch
• Avg.Dice coefficient : 0.973.

Apeer Workflow

• APEER platform to sum up was helpful in terms of the following usages,
  • Version-Control
  • Modular development
  • Easy Integration using Python

Our Apeer Workflow

Results

Input Image

Output Image