AI lesion tracking in PET/CT imaging: a proposal for a Siamese-based CNN pipeline applied to PSMA PET/CT scans

Area of Science:

• Nuclear Medicine
• Artificial Intelligence in Oncology
• Medical Imaging Analysis

Background:

• Assessing tumor response to systemic therapy is crucial in oncology, primarily using PET/CT scans.
• Current methods often analyze only a subset of lesions, risking bias due to inter-metastatic heterogeneity.
• Automated lesion tracking using AI can enhance the analysis of numerous lesions for a more accurate response assessment.

Purpose of the Study:

• To introduce a Siamese Convolutional Neural Network (CNN) approach for automated lesion tracking in PET/CT scans.
• To enable the analysis of a higher number of lesions for improved tumor response evaluation.
• To address the limitations of operator-dependent lesion selection in therapy monitoring.

Main Methods:

• The study applied a Siamese CNN to track bone lesions in full-body PSMA PET/CT scans of metastatic castration-resistant prostate cancer patients undergoing [177Lu]Lu-PSMA therapy.
• Lesion segmentation and affine registration were performed as data preparation steps.
• The algorithm extracted lesion patches and used a Siamese CNN trained to distinguish corresponding from non-corresponding lesion pairs.

Main Results:

• The Siamese CNN model demonstrated proficiency in classifying lesion assignments, achieving 83% accuracy and an AUC of 0.91 in its optimal configuration.
• The lesion tracking pipeline successfully identified corresponding lesions with an accuracy of 89%.
• Experiments explored various input patch types and both 2D and 3D CNN architectures.

Conclusions:

• A CNN-based approach can effectively facilitate the tracking of multiple lesions in PSMA PET/CT scans.
• This AI-driven method shows potential for improving the assessment of tumor response to systemic therapies.
• Further clinical studies are warranted to confirm if this enhanced lesion tracking improves therapy outcome prediction.