Deep Learning for Variational Multimodality Tumor Segmentation in PET/CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Positron emission tomography/computed tomography (PET/CT) provides both functional and anatomical data.
• Accurate tumor segmentation using fused PET/CT information remains a significant challenge.
• Existing methods struggle with complementary data fusion for precise delineation.

Purpose of the Study:

• To develop a novel deep learning-based variational method for automated multimodality tumor segmentation in PET/CT.
• To effectively fuse complementary functional (PET) and anatomical (CT) information.
• To improve the accuracy of tumor segmentation, especially in cases with challenging imaging features.

Main Methods:

• A 3D fully convolutional network (FCN) was trained to generate a tumor probability map from CT images.
• A fuzzy variational model incorporated the probability map and PET intensity for segmentation.
• A split Bregman algorithm was employed to optimize the variational model.
• The method was validated on a dataset of 84 non-small cell lung cancer PET/CT scans.

Main Results:

• The deep learning model required minimal training samples and is suitable for small clinical datasets.
• The proposed method outperformed existing deep learning and traditional multimodality segmentation techniques.
• Achieved high performance metrics: Dice Similarity Index (DSI) of 0.86 ± 0.05, Sensitivity (SE) of 0.86 ± 0.07, Positive Predictive Value (PPV) of 0.87 ± 0.10.
• Successfully segmented tumors with FDG uptake inhomogeneity, blurred edges, and complex surrounding tissues.

Conclusions:

• The novel deep learning variational method effectively fuses PET/CT data for accurate tumor segmentation.
• The approach demonstrates robustness in handling challenging imaging scenarios and small datasets.
• This method offers a significant advancement for automated tumor delineation in clinical practice.