A Comparative Study of Decoders for Liver and Tumor Segmentation Using a Self-ONN-Based Cascaded Framework

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate liver and tumor segmentation is vital for early liver cancer diagnosis.
• Manual interpretation is time-consuming and prone to errors.
• Computer-aided diagnosis (CADx) systems offer efficient alternatives.

Purpose of the Study:

• To develop and evaluate a cascaded encoder-decoder Self-Organized Neural Network (Self-ONN) for precise liver and tumor segmentation.
• To compare Self-ONN performance against other architectures like U-Net, FPN, and U-Net++.
• To assess the utility of transfer learning with ResNet and DenseNet backbones.

Main Methods:

• A cascaded encoder-decoder Self-ONN model was proposed for liver and tumor segmentation.
• The first stage segments the liver; the second stage segments tumors within the liver region of interest.
• Alternative architectures (U-Net, FPN, U-Net++) with ResNet/DenseNet backbones were investigated for comparison.

Main Results:

• Self-ONN with a ResNet18 backbone achieved 98.18% Dice Similarity Coefficient (DSC) and 97.44% Intersection over Union (IoU) for liver segmentation.
• Self-ONN with a DenseNet201 encoder achieved 92.84% DSC and 91.75% IoU for tumor segmentation.
• The proposed method demonstrated high accuracy in segmenting tumors of various sizes, including small lesions.

Conclusions:

• The cascaded Self-ONN model accurately segments liver tumors, outperforming manual methods.
• The system effectively identifies small and challenging-to-detect liver lesions.
• This AI-driven approach enhances early diagnosis of liver malignancies.