1. Abstract

2. Architecture

2.1 Overall Architecture

2.2 Architecture Specifications

2.3 Building Blocks

CompSeg (CSeg) Block

• Enhances MBConv efficiency through the Zoom-Filter-Rescale (ZFR) strategy.
• Involves:
  • Large stride at the entry expansion layer (Zooming).
  • Large kernel-sized depthwise convolution operation (Zooming and Filtering).
  • Rescaling spatial size using bilinear upsampling to compensate for downsampling (Rescaling).
• Increases receptive fields within MBConv by 8×.
• Decreases inference latency by approximately 30%.

Extended CompSeg (ECSeg) Block

• Extends the CSeg block by incorporating an improved global context (IGC) block for robust global context modeling.
• Increases AJI score of the nuclei segmentation on MoNuSeg 2018 dataset in a lightweight network by approximately 1.9%.
• Achieves improved performance with only a minimal computational cost.

Residual Bottleneck Transformer (RBT) Block

• Incorporates Transformer encoder blocks in a tailored manner to a variant of the Sandglass block.
• Leverages strengths of both convolutional and attention-based mechanisms.
• Effectively captures fine-grained spatial information and comprehensive long-range dependencies.

Noise-aware stem (Block) Block

• Stem block
• Enhances low-level feature extraction while mitigating the impact of noisy features in the model's initial skip connections.

3. Implementation

For cloning the project run:

   $ git clone https://github.com/mltraore/compsegnet.git   $cd compsegnet

Install requirements using:

   $ chmod +x install_dependencies.sh   $ pip install -r requirements.txt   $ sudo ./install_dependencies.sh

The following datasets were used in experiments:

• MoNuSeg 2018
• CoNSeP
• CPM-17

Processed versions of these datasets, along with the original versions, predicted masks, and a checkpoint sample for the MoNuSeg 2018 dataset, can be downloadedhere. For automatic download, run:

   $ chmod +x download_resources.sh   $ ./download_resources.sh

in the project directory.

• Note: The included predicted masks are actual CompSegNet architecture outputs, provided for qualitative comparison and further evaluation by the authors in their research.

To create the datasets yourself, use theprepare_dataset.py script:

# Get script usage info$ python3 prepare_dataset.py --help# Example usage$ python3 prepare_dataset.py --dataset-path datasets/monuseg_2018 \                             --image-size 1000 \                             --validation-size 0.10 \                             --reference-image datasets/monuseg_2018/train/images/1.tif \                             --prepared-data-path datasets/prepared_datasets/monuseg_2018

Use thetrain.py script in the project directory to train the model:

# Get script usage info$ python3 train.py --help# Example usage$ python3 train.py --train-folder datasets/prepared_datasets/monuseg_2018/train \                   --validation-folder datasets/prepared_datasets/monuseg_2018/validation \                   --checkpoints-folder checkpoints/ckpts

Use thetest.py script in the project directory to test the model:

# Get script usage info$ python3 test.py --help# Example usage$ python3 test.py --model-weights-save checkpoints/ckpts \                  --test-set datasets/prepared_datasets/monuseg_2018/train

This provides Dice Similarity Coefficient (DSC), Aggregated Jaccard Index (AJI), and Hausdorff Distance (HD) scores.

4. Results

4.1 Quantitative Results

4.2 Qualitative Results

Citation

Acknowledgement