segment mitochondria in microscope images.
There is 141 images, each has a resolution of 1024*768 pixels. I use 126 of them as training set, and leave the rest for testing. For each picture, I extract 64*64 patch, then apply data augmentation to get more training samples. With stride of 64, I had 192 patches from each picture, then by rotating 3 times, and applying horizontal and vertical mirroring, I can get 6 times more, which is 1152. Hence the total amount of training set is 145152.
The model is almost U-Net, one big difference is, pooling layers and transposed convolution layers have both been reduced to 3 times, different from the original version, which has it 4 times. And the number of channels of all inner layer has been halved. please read the source code for the details.
I use opening and closing to reduce small noises. Opening is defined as erosion followed by dilation, while closing is the opposite, defined as dilation then erosion. I found that doing several times of closing then opening would make the result a bit better.
Training is performed on Google Colaboratory. Batch size is fixed at 64, after about 500 epochs, it reaches probably the best result. mito.ipynb shows the process.
Thresholded output of the model is in pred_6, I deployed post processing on that output, and the result is in rst.
I assume mitochondria as positive, and other area is negative. Below is probably the best result I did.
| Accuracy | Sensitivity | Specificity | Dice | IOU |
|---|---|---|---|---|
| 0.9885 | 0.9606 | 0.9901 | 0.8984 | 0.8156 |