Deep Learning-powered CT-less Organ Segmentation from 68Ga-PSMA/Dotatate PET Images

• Main Authors: Salimi, Y., Mansouri, Z., Shiri, I., Zaidi, H.
• Format: Conference Proceeding
• Language: English
• Subjects: Computed tomography; Image segmentation; Pipelines; Prediction algorithms; Reliability; Semiconductor detectors; Spleen
• ISSN: 2577-0829
• DOI: 10.1109/NSS/MIC/RTSD57108.2024.10655415

Organ segmentation and quantification can help to improve the diagnostic and prognostic value of PET/CT imaging. The use of widely available CT-based organ segmentation approaches inherently suffers from limitations owing to the high prevalence of mismatch between CT and PET images, mainly due to respiratory phase differences. This study aimed to develop a CT-less organ segmentation pipeline using only emission Ga-68 PSMA/Dotatete PET images. A total number of 570 patients injected with Ga-68-PSMA or Ga-68-Dotatate were included in this study. After excluding images presenting with mismatch, CT, or PET artifacts, 185 PET/CT images were retained. A total number of 14 organs were delineated automatically employing previously developed deep learning algorithms in our lab for CT image segmentation. Subsequently, the segmentation masks were resampled to PET resolution. The PET non-corrected and segmentation masks were used to train self-configuring nnU-Net models in five-fold data split strategy using 2000 epochs. Dice and Jaccard metrics were used to compare the predicted and reference segmentations masks. Average Dice values of 0.82 \pm 0.04,0.93 \pm 0.03,0.69 \pm 0.12,0.86 \pm 0.06, 0.81 \pm 0.04,0.83 \pm 0.09,0.87 \pm 0.08,0.90 \pm 0.03,0.62 \pm 0.12,0.56 \pm 0.07,0.81 \pm 0.04,0.81 \pm 0.10,0.82 \pm 0.09,0.82 \pm 0.02 was achieved for the aorta, brain, eyeballs, heart, hips, kidneys, liver, lungs, pancreas, ribs, sacrum, spleen, urinary bladder, and vertebrae, respectively. Ga-68 PET image quantification using segmentation masks generated on CT images is limited to available and reliable co-registered CT images which is not always the case due to mismatch between PET and CT images. We developed a robust and fast organ segmentation pipeline to tackle this issue using non-corrected PET images.