Single-Stage Lesion Identification in $^{68}$Ga-DOTATATE PET Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Positron emission tomography (PET) is crucial for neuroendocrine tumor (NET) assessment.
• Automated lesion identification in PET images is vital for effective NET treatment development.
• Existing deep learning methods often require predefined regions of interest (ROIs/VOIs) or complex multi-stage pipelines, leading to inefficiency and variability.

Purpose of the Study:

• To propose a novel single-stage PET lesion detection method for NETs.
• To eliminate the need for precomputed ROIs/VOIs, cascaded models, or multimodal data.
• To enhance the efficiency and accuracy of lesion identification in PET imaging.

Main Methods:

• Introduction of a three-dimensional dual-decoder neural network.
• Incorporation of a cross-decoder attention module using auxiliary organ segmentation.
• Integration of a patchwise contrastive learning module to improve lesion discriminative power.

Main Results:

• The proposed method was evaluated on hepatic NET $^{68}$Ga-DOTATATE PET image datasets from two different scanners.
• Superior performance was achieved compared to reference baselines and state-of-the-art approaches.
• The single-stage framework demonstrated significant improvements in lesion identification.

Conclusions:

• The novel single-stage framework, utilizing cross-decoder attention and patchwise contrastive learning, effectively improves lesion identification in PET images.
• This approach has the potential to significantly enhance the efficiency of clinical interpretation of PET imaging data for NETs.