Liver and Liver Tumor Segmentation using W-Net
W-Net is a novel 2D CNN that we developed for liver and liver tumor segmentation. The proposed model achieved highest ever recorded dice score on the LiTS-17 dataset.
Created on 20th March 2023
•
Liver and Liver Tumor Segmentation using W-Net
W-Net is a novel 2D CNN that we developed for liver and liver tumor segmentation. The proposed model achieved highest ever recorded dice score on the LiTS-17 dataset.
The problem Liver and Liver Tumor Segmentation using W-Net solves
Liver tumor segmentation is regarded as one of the most difficult semantic segmentation tasks in medical imaging. High variability in the imaging platforms and the location of the tumor within the organ makes this task especially challenging. We have proposed a unique preprocessing scheme that extracts 2D slices from the DICOM volumes prior to training. Unlike traditional 3D liver-tumor segmentation models ours is based on a 2D framework which greatly reduces the computational load. the model also uses two encoders- one with pre-trained weights and the other without, giving it a W shape. The model has achieved the high ever recorded dice scores on the LiTS-17 dataset making it a clinical standard.
Challenges I ran into
The main Idea here is to develop an application for liver and liver-tumor segmentation which provides outputs that are acceptable by clinical standards. The central backbone of the project is the preprocessing methodology and the model architecture. The preprocessing methodology entails taking 2D slices out of the DICOM volumes, applying HU windowing to those slices, and then applying a contrast-limited adaptive histogram equalisation filter. The preprocessed slices and corresponding masks were fed to the model for training. The proposed model architecture is based on the encoder-decoder paradigm. It uses two encoders and one decoder. One of the encoders has a ResNet-50 backbone pre-trained on the imagenet dataset and the other has no pretrained weights. The proposed approach yielded the highest ever recorded dice scores on liver and liver tumor segmentation on the LiTS-17 dataset which is a clinical standard for liver tumor segmentation.
Technologies used