Learning Without Real Data Annotations to Detect Hepatic Lesions in PET Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep neural networks (DNNs) for lesion identification in PET imaging typically require extensive annotated data.
• Acquiring sufficient annotated PET data for rare diseases like neuroendocrine tumors (NETs) is challenging due to low incidence and high annotation costs.

Purpose of the Study:

• To develop an adaptable deep learning framework for hepatic lesion detection in NETs using low-cost, simulated data instead of real lesion annotations.
• To improve the generalizability and reduce annotation burden for lesion detection in clinical PET imaging.

Main Methods:

• Proposed a region-guided generative adversarial network (RG-GAN) for lesion-preserved image-to-image translation.
• Developed a data augmentation module specifically for list-mode simulated data to enhance model training.
• Integrated RG-GAN, data augmentation, and a lesion detection network into a unified framework for joint-task learning.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art lesion detection techniques in real clinical 68Ga-DOTATATE PET images.
• Achieved performance comparable to models trained with actual lesion annotations, validating the effectiveness of simulated data.

Conclusions:

• Effective hepatic lesion detection in NETs can be achieved without real data annotations by utilizing RG-GAN modeling and specialized data augmentation.
• This adaptable deep learning approach significantly reduces the effort required for data annotation and enhances model generalizability for PET lesion detection.