Prediction of Disease Quantification via PET/CT Images and Auto-Segmentation of Target Objects

In this paper, we propose a deep learning-based disease quantification (DQ) method which predicts object-level total lesion burden (TLB) of target objects from segmentations derived by automatic methods on PET/CT images. Our approach employs a U-Net shaped network using CT images, standardized SUV images, and delineated object mask as three-channel input, and predicted segmentation mask and lesion map (LM) as two-channel output. Ground truth LM is generated based on normality distribution of metabolic activities of the object, and the loss function is a weighted sum of segmentation loss, LM loss, and TLB loss. We evaluated the method on clinical datasets comprising 58 PET/CT scans, covering five organs and five lymph node zones. 3-fold cross-validation experiments were conducted using both manual and auto-segmented masks. Results show that the proposed method achieves clinically acceptable errors, where half of experiments showed errors below 15% and most of the remainder with errors under 30%. Our method demonstrates a solution for automated DQ from PET/CT images, potentially reducing manual effort while maintaining clinically meaningful accuracy and reproducibility.