An Automated Deep Learning-Based Framework for Uptake Segmentation and Classification on PSMA PET/CT Imaging of...

Area of Science:

• Nuclear Medicine
• Radiology
• Artificial Intelligence

Background:

• Accurate segmentation and classification of uptake on prostate-specific membrane antigen (PSMA) PET/CT scans are crucial for automating whole-body tumor burden assessments in prostate cancer.
• Current methods may lack the precision required for reliable automated analysis.

Purpose of the Study:

• To develop and evaluate an automated deep learning (DL) framework for segmenting and classifying uptake on PSMA PET/CT scans.
• To improve the accuracy of lesion identification and characterization for better tumor burden determination.

Main Methods:

• A DL-based framework using two Convolutional Neural Networks (CNNs) was developed for segmentation and classification.
• Anatomical prior guidance was integrated to focus the DL framework on PSMA-avid lesions.
• The framework was trained and tested on 193 [18F]DCFPyL PET/CT scans from two institutions, with segmentation and classification performance evaluated using metrics like Dice, IoU, precision, recall, accuracy, and AUC.

Main Results:

• The DL-based segmentation, guided by anatomical priors, showed improved performance with mean Dice of 0.700 (internal) and 0.680 (external).
• The multi-modal decision fusion classification framework achieved high accuracy (0.764 internal, 0.796 external) and AUC (0.863 internal, 0.851 external) in distinguishing suspicious from non-suspicious foci.
• The framework demonstrated robust performance on both internal and external test sets, outperforming single-modal and multi-modal CNNs.

Conclusions:

• DL-based lesion segmentation on PSMA PET/CT is effectively facilitated by the proposed anatomical prior guidance strategy.
• The developed classification framework accurately differentiates suspicious foci from those not suspicious for cancer, offering a valuable tool for automated tumor burden assessment.