The challenge in this project was to detect functional tissue units (FTUs) across different tissue preparation pipelines. An FTU is defined as a “three-dimensional block of cells centered around a capillary, such that each cell in this block is within diffusion distance from any other cell in the same block” (de Bono, 2013). The goal is the implementation of a successful and robust glomeruli FTU detector.
This is a semantic segmentation task to identify glomeruli in a given kidney tissue image. This is done as a part of a Kaggle competition. The provided images were really large images with average size is equal to 25 x 30 (k). Such image does not fit in the memory for processing. However one image contained a large number of gromeruli and by tiling the large image in to small set of images created a reasonably managable and representative training and testing image data sets. Therefore the solution is mainly consist of two steps.
Example of an original image and its mask (31299 x 44066 pixels)
- Install a library called patchify to tile large images into small parts
- Make two folders to hold images and masks
- Make a list of tiff image file names in a given floder
- Read tiff files one by one (each file differ by number of channels and where the color chanel is located on)
- Use patchify to break the image to 4000 x 4000 tiles, do this by every 2000 pixels
- Read the corresponding mask for the image given by json specification file
- Generate mask tiles using patchify as per the same parameters
- For each image and mask tile calculate the percentage of the mask covered in the whole image
- Select all images where the mask is between 10-90% comapred the whole image
- Resize each image and mask to 400 x 400
- Rename those in a way that it can be identified uniquely using the row and column in which those are located in the original image
- Save the resized image and mask in the corresponding folder
Example of image and its corresponding mask saved in the disk
- Read image and mask file names into a list and shuffle it
- Create training and test images list using above list
- Generate tensorflow datasets for both training and test images
- Define custom evaluation matrics calle dice coefficient and dice loss
- Install segmentation models library
- Define a Unet model with a desired backbone and pre loaded weights
- Test the model with the test dataset
- Train the Unet model and save the model in h5 format
Example of a prediction in the test set using the saved model
From left to right, The original image, its orignal mask and the predicted mask using the trained Unnet model
The performance was evaluated on the mean Dice coefficient. The Dice coefficient can be used to compare the pixel-wise agreement between a predicted segmentation and its corresponding ground truth. The formula is given by:
where X is the predicted set of pixels and Y is the ground truth. The Dice coefficient is defined to be 1 when both X and Y are empty. The final score is the mean of the Dice coefficients for each image in the test set.
- Dice coefficient: 0.86