UNSUPERVISED ADAPTATION FROM FDG TO PSMA PET/CT FOR 3D LESION DETECTION UNDER LABEL SHIFT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Unsupervised domain adaptation (UDA) is crucial for adapting medical imaging models across different datasets.
• Cross-tracer adaptation in PET/CT (e.g., FDG to PSMA) presents challenges beyond covariate shift, including label shift in lesion characteristics.

Purpose of the Study:

• To develop a UDA framework for 3D volumetric lesion detection that adapts detectors from labeled FDG PET/CT to unlabeled PSMA PET/CT data.
• To explicitly model and compensate for label shift in lesion size and count during cross-tracer adaptation.

Main Methods:

• Proposed a self-training framework with two label-shift compensation mechanisms: adaptive anchor shape adjustment and size bin-wise pseudo-label quotas.
• Employed an exponential moving average for anchor updates and a histogram-based approach for pseudo-label selection.
• Alternated between supervised learning on labeled FDG data and pseudo-labeled PSMA data.

Main Results:

• The proposed method significantly improved Average Precision (AP) and Free-response Receiver Operating Characteristic (FROC) performance compared to source-only and conventional self-training baselines.
• Demonstrated effectiveness on the AutoPET 2024 dataset, adapting from 501 FDG studies to 369 PSMA studies.
• Showcased that modeling target lesion prevalence and size composition enhances cross-tracer detection robustness.

Conclusions:

• The developed UDA framework effectively addresses label shift in cross-tracer PET/CT lesion detection.
• Explicitly modeling target domain lesion characteristics is key for robust performance in unsupervised adaptation scenarios.
• This approach offers a promising direction for improving 3D volumetric lesion detection in medical imaging.