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

Positron emission tomography (PET) is a commonly used imaging modality for assessment of neuroendocrine tumors (NETs), and lesion identification in PET images is a key step in the development of effective treatments. Deep neural networks have recently produced encouraging performance of automated lesion identification with PET imaging. However, most methods require a predefined region/volume of interest (ROI/VOI) or rely on a multi-stage, cascaded modeling pipeline, which often leads to low efficiency and/or high variability. In this paper, we propose a novel single-stage PET lesion detection method that does not need precomputed ROIs/VOIs, cascaded models or multimodal data.

We introduce a novel three-dimensional dual-decoder neural network, which contains a cross-decoder attention module to take as input gating signals from an auxiliary organ segmentation decoder and suppress irrelevant feature responses in the primary decoder of lesion detection. Additionally, we design and insert a new patchwise contrastive learning module into the primary decoder to enhance the network's discriminative power for lesions with varying volumes and shapes.

We evaluate the proposed lesion identification method using multiple hepatic NET $^{68}$Ga-DOTATATE PET image datasets that are acquired from two different scanners. The method produces superior performance compared with the reference baseline and recent state-of-the-art approaches.

We propose a novel single-stage framework, where both the cross-decoder attention and the patchwise contrastive learning are beneficial to improvement of lesion identification performance in PET images.

The proposed study has the potential to significantly improve the efficiency of clinical interpretation of PET imaging data.