Multi-class Tissue Segmentation of CT images using an Ensemble Deep Learning method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Microwave ablation (MWA) therapy uses CT images for lung tumor destruction.
• Tumor recurrence due to insufficient ablation necessitates precise tumor area determination.
• Current deep learning segmentation methods require large datasets, posing a challenge for MWA therapy.

Purpose of the Study:

• To develop a semantic segmentation method for accurate organ and tumor segmentation in CT images during MWA therapy.
• To overcome the limitation of small datasets in deep learning for MWA tumor segmentation.
• To improve the precision of lung cancer treatment by accurately identifying tumor boundaries.

Main Methods:

• Developed four U-Net based semantic segmentation models.
• Combined outputs from multiple models to enhance segmentation accuracy.
• Utilized a consensus approach by selecting the highest weight value for the final segmentation.
• Evaluated performance using Intersection over Union (IoU) for background, lung, ablated, and tumor tissues.

Main Results:

• Achieved high average IoU scores: 0.99 (background), 0.98 (lung), 0.77 (ablated), and 0.54 (tumor).
• Demonstrated superior performance of the multi-model approach compared to individual models, especially with small datasets.
• Successfully segmented lungs, lung tumors, and ablated tissues in CT medical images.

Conclusions:

• The proposed multi-model deep learning approach enhances semantic segmentation accuracy for MWA therapy.
• This method is effective even with limited medical imaging data, addressing a key challenge in clinical practice.
• Improved AI-driven segmentation can aid clinicians in assessing complete tumor destruction, potentially reducing recurrence rates.