Monday, December 2 | 3:10 p.m.-3:20 p.m. | M7-SSNMMI03-2 | Room S405
In this session, researchers will present an automated deep-learning AI model that could help clinicians segment rare neuroendocrine tumors (NETs) on whole-body 3D gallium-68 (Ga-68) DOTATATE PET/CT images.
Currently, segmentation of tumor lesions is performed manually, which is tedious and time-consuming, the group noted. Presenter Fahmida Haque, PhD, a postdoctoral research fellow, and colleagues at the National Cancer Institute, specifically developed the model for use in imaging pheochromocytoma and paraganglioma (PHEOs/PGLs), a rare NET that can lead to extensive metastases in the whole body and can be surgically inoperable. They used 129 whole-body Ga-68 DOTATATE PET/CT scans from 38 patients with inoperable PHEOs/PGLs and split them into groups.
The researchers trained a 3D full-resolution nnUNet model (3D_FullRes) on 55 of the images with fivefold cross-validation. The model's performance was evaluated on a test set of 60 images for lesion segmentation using the dice similarity coefficient (DSC). In addition, to validate the diagnostic performance of the model, they conducted both lesion-level and patient-level statistical analysis, including total lesion glycolysis (TLG) and metabolic tumor volume (MTV).
The model achieved a DSC of 0.87, a sensitivity of 87%, and a positive predictive value of 88%, with four median false-positive lesions per scan on the test set, according to the results. The average mean difference with 95% confidence interval and Spearman correlation coefficient between AI-predicted and the MTV and TLG of ground truth (GT) lesions were -32.27 (-281.06, 216.53), -395.44 (6072.86, -6863.73), and 0.95 and 0.96, respectively, which was slightly less than ground truth annotations because of false negatives. An anatomical position-based failure analysis showed that most of the false negative lesions (70 out of 504 lesions) were observed in the liver, the group noted.
“The proposed lesion-segmentation model for DOTATATE PET/CT scan can help imaging physicians in lesion segmentation, tracking progression, and determining better treatment decisions for patients with PHEOs/PGLs,” the group concluded.
Tune in Monday afternoon in this scientific session on neuroendocrine tumors to learn all the details.
