Tumor co-segmentation in PET/CT using multi-modality fully convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate tumor segmentation is crucial for cancer diagnosis and treatment planning.
• Current deep learning methods often lack multi-modality integration, limiting their effectiveness.
• Positron Emission Tomography/Computed Tomography (PET/CT) offers combined metabolic and anatomical imaging data.

Purpose of the Study:

• To develop a novel multi-modality deep learning segmentation method for tumors using PET/CT data.
• To simultaneously leverage metabolic (PET) and anatomical (CT) information for enhanced segmentation accuracy.
• To evaluate the proposed method's performance against existing techniques on a clinical lung cancer dataset.

Main Methods:

• A 3D fully convolutional neural network (FCN) architecture was designed for simultaneous PET and CT processing.
• The network employed parallel sub-segmentation CNNs for feature extraction from each modality.
• A feature fusion module with cascaded convolutional blocks integrated features using weighted cross-entropy minimization.

Main Results:

• The proposed method demonstrated significant performance gains compared to single-modality and existing co-segmentation techniques.
• Validation on a clinical dataset of 84 lung cancer patients confirmed the method's effectiveness, speed, and robustness.
• The approach outperformed traditional methods, V-net based co-segmentation, fuzzy set theory methods, and a W-net deep learning model.

Conclusions:

• The novel 3D FCN effectively integrates PET and CT data for accurate tumor segmentation.
• This multi-modality approach offers a significant advancement over single-modality deep learning methods.
• The proposed method shows promise for improving computer-aided diagnosis and treatment in lung cancer.