Biomedical Cell Segmentation Model Comparison

Abstract: Image segmentation has long been used for a variety of applications in biomedical imaging and classification. Particularly useful is its ability to identify key regions of interests in cellular contexts, specifically in cell structures and organelles. By analyzing various U-Net architectures on a challenging dataset of cellular images and corresponding masks, we can assess various architectures and identify optimal model parameters and layers that allow for optimal segmentation of key regions of interests. Such experiments can help guide approaches in more complex segmentation tasks utilizing deep learning models. This study not only aids in model selection for practitioners but also sets the stage for further refinement and real-world application of advanced segmentation techniques in biomedicine. Specifically, I explore the efficacy and accuracy of 2 U-Net models that leverage slightly different architectures and attention schemes on the 2018 Kaggle Data Science Bowl Dataset, a collection of 735 high quality cellular images and corresponding masks.

Project Github

U-Net Overview

Purpose: U-Net is a convolutional neural network originally developed for biomedical image segmentation.

Architecture: Incorporates an encoder which captures context and a symmetrically adjacent decoder which enables precise localization.

Strengths:

- Effective with limited data.

- Applicable in a variety of medical imaging contexts (cell segmentation, tumor detection, organ segmentation, etc.)

Image From: Towards Data Science

Attention U-Net Overview

Purpose: U-Net is a convolutional neural network originally developed for biomedical image segmentation.

Architecture: Introduces attention gates at the skip connections, which helps the model focus on target structures and features.

Strengths:

- Improved Accuracy

Image From: Research Gate

DataSet + Metrics

2018 Kaggle Data Science Bowl Dataset

- 670 Training Images

- 536 Training

- 134 Validation

- 65 Test Images

Dice Coefficient

- Statistical Measure evaluating the similarity between 2 sets.

- The value of the Dice coefficient ranges from 0 to 1.

- 0 indicates no overlap and 1 indicates perfect overlap between the predicted and ground truth masks.

Experiments

Experiment 1: U-Net

For U-Net, I did a 70-15-15 split of the dataset into train, validation, and test data, respectively.

The input image is of size 128 by 128, with 32 filters.

I used the Adam optimizer and binary cross entropy as the loss function.

I trained for 20 epochs with a batch size of 32.

More information on the details of the network can be found in the attached Github Repository

While training, the model got a maximum training accuracy of 0.9702 and a maximum validation accuracy of 0.9711.

On the test set, the model performed just as well, with an accuracy of 0.9710.

Experiment 1 Sample Performance

Experiment 2: Attention U-Net

For Attention U-Net, we train with similar input parameters (including an input image size of 128 by 128). For more details on the code and architecture, feel free to check out my project's github repository.

While training, we get a maximum training accuracy of 0.8993 and a maximum validation accuracy 0.8960.

On the test set, we get an accuracy of 0.8549.

Experiment 2 Sample Performance

Results

You may also like