A dual-encoder double concatenation Y-shape network for precise volumetric liver and lesion segmentation
Gabriella d'Albenzio1, Yuliia Kamkova2, Rabia Naseem3
1The Intervention Center, Oslo University Hospital, 0slo, Norway; Department of Informatics, University of Oslo, Oslo, Norway.
Computers in Biology and Medicine
|July 18, 2024
Summary
A new Dual-Encoder Double Concatenation Network (DEDC-Net) accurately segments livers and tumors in CT scans. This deep learning approach improves hepatocellular carcinoma diagnosis and surgical planning by overcoming segmentation challenges.
Area of Science:
- Medical Imaging
- Deep Learning
- Computational Anatomy
Background:
- Accurate segmentation of liver and tumors in CT scans is vital for hepatocellular carcinoma (HCC) diagnosis and surgical planning.
- Deep learning methods face challenges in automated abdominal CT segmentation due to class imbalance and structural variations.
- Existing cascaded approaches for segmentation are computationally expensive.
Purpose of the Study:
- To introduce the Dual-Encoder Double Concatenation Network (DEDC-Net) for simultaneous liver and tumor segmentation.
- To enhance feature reuse and optimize segmentation performance using residual and skip connections.
- To provide a computationally efficient and accurate deep learning solution for liver and tumor segmentation.
Main Methods:
- Developed DEDC-Net, a novel deep learning architecture utilizing dual encoders with residual and skip connections.
- Evaluated DEDC-Net on the LiTS dataset for liver and tumor segmentation.
- Conducted ablation studies comparing VGG19 and ResNet backbones and assessing the impact of attention mechanisms.
- Validated the model's robustness on unseen CT datasets (IDCARDb-01 and COMET).
Main Results:
- DEDC-Net achieved a mean Dice Score (DS) of 0.898 for liver segmentation on the LiTS dataset, outperforming existing methods.
- Integrating residual connections in one encoder maximized tumor segmentation DS.
- The model demonstrated superior lesion segmentation on the IRCADb-01 dataset with a DS of 0.629.
- DEDC-Net achieved state-of-the-art performance without additional attention gates.
Conclusions:
- DEDC-Net offers an effective and efficient solution for simultaneous liver and tumor segmentation in CT volumes.
- The proposed network architecture enhances feature representation and improves segmentation accuracy.
- The model's performance on multiple datasets highlights its generalizability and robustness for clinical applications.
