Location-Aware Encoding for Lesion Detection in ⁶⁸Ga-DOTATATE Positron Emission Tomography Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lesion detection in Positron Emission Tomography (PET) is crucial for staging, treatment planning, and evaluating novel therapies, particularly for neuroendocrine tumors (NETs).
• Current methods often involve manual region of interest (ROI) cropping, multi-stage models, or multi-modal imaging, leading to inefficiencies and potential errors.
• There is a need for more efficient and accurate automated lesion detection in PET imaging.

Purpose of the Study:

• To propose a novel, single-stage deep learning method for lesion detection using only PET images.
• To enhance lesion location-specific feature learning at multiple scales.
• To improve the efficiency and accuracy of lesion detection in PET imaging for NETs.

Main Methods:

• A U-Net-like neural network architecture was designed with a plug-and-play codebook learning module.
• The network incorporates multi-level supervision for codebook learning to promote discriminative feature extraction at multiple scales.
• A learnable fusion layer automatically combines predictions from the codebook module and other network layers.

Main Results:

• The proposed method was evaluated on a clinical ⁶⁸Ga-DOTATATE PET image dataset.
• The single-stage method demonstrated significantly superior lesion detection performance compared to existing state-of-the-art approaches.
• The approach eliminates the need for manual ROI/VOI cropping, multi-stage modeling, and multi-modality data.

Conclusions:

• A novel deep learning method for efficient, single-stage lesion detection in PET imaging has been presented.
• This method offers a new perspective for effective lesion identification, potentially accelerating therapeutic development for NETs.
• The findings may lead to improved patient outcomes and survival rates through enhanced diagnostic capabilities.