Semantic Segmentation models have been trained on the MoNuSeg Dataset comprising of 30 H&E stained tissue training Images of 30 patients with tumors of different organs. The dataset also has a subset of 14 images for validation of the models. The images were taken at 40 times magnification.
https://monuseg.grand-challenge.org/Data/
Three architectures have been used for learning purposes:
- Segnet
- Unet
- DeepLabV3
The implementation and results of each have been provided in their respective Jupyter notebooks in this repository.
The metrics that have been used are Loss, Accuracy, Dice Coefficient and F1-Score:
| Modified Unet | DeepLabV3 | Segnet | |
|---|---|---|---|
| Loss | 0.3436 | 0.4292 | 0.6813 |
| Accuracy | 91.40% | 90.59% | 83.26% |
| Dice | 0.7728 | 0.7589 | 0.3226 |
| F1 | 0.7889 | 0.7766 | 0.5792 |
The The best results have been obtained with the Modified Unet
The performance plots for the models are:
For improving the accuracy of the model, multiple techniques were applied in order to get good results with good training and inference times. With the following changes the performance slightly increased and the training time significantly reduced (50 Epocs in only 6 minutes on Google Colab).
By doing so the network has simplified and it's speed has dramatically increased.
By doing this the performance of the network slightly increased.
Removed the middle dense layer from SE Block that decreases (normally by a ratio of 16) the number of neurons that are input to the final dense layer of SE block for generating excitation values.
The settings for Unet that were used are given below:
- Optimizer = Adam
- Initial Learning Rate = 0.001
- Learning Rate Decay by factor of 0.1 if no improvement is seen
- Loss = Binary Crossentropy
- Initial Number of filters for Unet = 16
- Drop Out = 0.05
- Batch Size = 16
10% of the training set was used as validation set during training of the network.
Actual image size is 1000x1000 but for training purposes patching of 256x256 has been implemented.
The accuracy of the Unet is slightly greater than DeepLabV3 and it is very light weight compared to DeepLab.
Following are the curves for accuracy, dice coefficient and f1-score for the Unet model:
