CT Liver Segmentation Via PVT-based Encoding and Refined Decoding

Overview

PVTFormer is a novel encoder-decoder framework designed for precise liver segmentation from CT scans. At its core, it utilizes the Pyramid Vision Transformer (PVT v2) as a pretrained encoder, enhancing the segmentation process with its unique ability to handle variable-resolution input images and produce multi-scale representations. Our approach includes a novel hierarchical decoding strategy that incorporates specialized upscaling in the Up block with effective multi-scale feature fusion in the Decoder. This approach significantly enhances the network's ability to delineate detailed semantic features, which is vital for precise liver segmentation.

Key Features

-Encoder-Decoder Framework: Incorporates PVT v2 for efficient and rich feature extraction.

-Hierarchical Decoding Strategy: Enhances semantic features for high-quality segmentation masks.

-Efficient Feature Processing: Combines residual learning with Transformer mechanisms for optimal feature representation.

-High Performance Metrics: Achieves impressive dice coefficients and mean IoUs, outperforming state-of-the-art methods.

PVTFormer Architecture

Architecture Advantages:

Improved accuracy for liver and other medical image segmentation.
Efficient learning of hierarchical features.
Ability to capture long-range spatial dependencies.

Applications

PVTFormer is highly effective for healthy liver segmentation, with potential applications in other medical imaging areas. It represents a significant advancement in medical image segmentation, offering a robust solution for accurate diagnosis and treatment planning.

Uses of PVTFormer:

Medical Image Segmentation
General Image Segmentation
Anomaly Detection in Medical Images
Comparative Studies

Dataset

LiTS dataset

Results

Qualitative results comparison of the SOTA methods

Quantitative results comparison of the SOTA methods

Citation

Please cite our paper if you find the work useful:

@inproceedings{jha2024ct,
  title={CT Liver Segmentation via PVT-based Encoding and Refined Decoding},
  author={Jha, Debesh and Tomar, Nikhil Kumar and Biswas, Koushik and Durak, Gorkem and Medetalibeyoglu, Alpay and Antalek, Matthew and Velichko, Yury and Ladner, Daniela and Borhani, Amir and Bagci, Ulas},
  bookarticle={Proceedings of the International Symposium on Biomedical Imaging (ISBI)},
  year={2024}
}

Contact

Please contact debesh.jha@northwestern.edu for any further questions.

DebeshJha / PVTFormer

readme