Liver Tumors Semantic Segmentation From CT Scans Using Polar Transformations

Overview

Focus of the project is on the impact of polar transformations on the performance of liver tumors semantic segmentation using Deep Convolutional Neural Networks, similar to [1].

Data

The data used to train, validate the models is from the LiTS dataset and the test data is from the 3DIRCADB dataset.

Models

For the purposes of this work, 3 popular semantic segmentation models have been considered: U-Net, U-Net++ with a ResNet encoder and DeepLabV3+ with a ResNet encoder.

Training

All models have been trained from scratch (no transfer learning has been done) using data augmentation techniques such as rotation, horizontal and vertical flip. The training configuration is defined in the './src/model_training/training_config.py' file. Training experiments have been handled, logged and monitored using the W&B service.

Each model has been trained in two separate settings: the carthesian setting and the polar setting, each of which are displayed in the diagrams below.

Carthesian Model Training Pipeline

Polar Model Training Pipeline

For the polar setting, the polar transformation was applied using as polar origin the center of mass of the biggest annotated blob in the corresponding mask for each CT scan slice.

Results

Carthesian Models

Model	Average Test Dice Score
U-Net	78.67%
U-Net++	75.48%
DeepLabV3+	80.92%

Polar Models - Test Set-up 1

Model	Average Test Dice Score
U-Net	64.12%
U-Net++	67.26%
DeepLabV3+	52.22%

Polar Models - Test Set-up 2

Model	Average Test Dice Score
U-Net	69.03%
U-Net++	79.67%
DeepLabV3+	59.99%

U-Net Sample Prediction

U-Net++ Sample Prediction

DeepLabV3+ Sample Prediction

References

• 1: "Training on Polar Image Transformations Improves Biomedical Image Segmentation"