Disease quantification on PET/CT images without explicit object delineation.

Area of Science:

• Medical Imaging
• Radiology
• Computational Pathology

Background:

• Image segmentation remains a significant challenge for quantitative analysis in clinical imaging.
• Accurate disease quantification from PET/CT scans is crucial for patient management.

Purpose of the Study:

• To present a novel approach, automatic anatomy recognition-disease quantification (AAR-DQ), for disease quantification on PET/CT images.
• To decouple disease quantification (DQ) from explicit organ delineation using automatic anatomy recognition (AAR).

Main Methods:

• The AAR-DQ method utilizes AAR to recognize anatomical objects and build optimal disease maps.
• Fuzzy modeling and fuzzy masks are employed to quantify disease burden (TLG, SUV) from PET/CT images, handling partial volume effects.
• Disease quantification can be performed automatically (DQ-AO) or semi-automatically (DQ-MO) at the organ level, or semi-automatically (DQ-ML) at the lesion level.

Main Results:

• Testing on 67 datasets (phantoms and patients) showed overall errors of approximately 6%, 3%, and 0% for TLG, SUVmean, and SUVmax, respectively.
• Errors were comparable to literature for large lesions and superior for smaller lesions, with reduced data and computational demands.
• DQ-MO generally provided more accurate results than DQ-AO, though not statistically significant.

Conclusions:

• Accurate disease quantification on PET/CT images is feasible using object recognition, bypassing explicit lesion delineation.
• The AAR-DQ approach, particularly DQ-AO and DQ-MO, holds potential for routine, body-wide disease burden quantification.
• This method offers an advancement over existing techniques by enabling organ-level quantification and handling smaller lesions more effectively.