Evaluating a Machine Learning Tool for the Classification of Pathological Uptake in Whole-Body PSMA-PET-CT Scans

Area of Science:

• Nuclear medicine imaging
• Machine learning applications in oncology
• Radiomics and computational pathology

Background:

• Differentiating pathological from physiological tracer uptake in PSMA-PET/CT scans is critical for prostate cancer diagnosis and treatment planning.
• Manual assessment of these images is time-consuming and requires significant attention.
• Supervised machine learning offers a potential solution to automate and improve this diagnostic process.

Purpose of the Study:

• To compare and validate supervised machine learning algorithms for classifying pathological uptake in prostate cancer (PC) patients.
• To assess the performance of machine learning in distinguishing between pathological and physiological tracer uptake on 68Ga-PSMA-PET/CT images.
• To evaluate the potential of machine learning as a tool to assist in clinical diagnosis.

Main Methods:

• Retrospective analysis of 68Ga-PSMA-PET/CT scans from 72 prostate cancer patients.
• Extraction of 77 radiomics features from 2452 manually delineated hotspots, labeled as pathological or physiological.
• Training and validation of three supervised machine learning classifiers using a held-out test dataset of 331 hotspots from 15 patients.

Main Results:

• High overall average performance with an area under the curve (AUC) of 0.98 was achieved.
• Excellent detection of pathological uptake with a mean sensitivity of 0.97.
• Moderate specificity (0.82) for detecting physiological uptake, with challenges noted for gland-specific uptake.

Conclusions:

• Supervised machine learning algorithms can predict hotspot labels with high accuracy on unseen PSMA-PET/CT data.
• The developed ML algorithm shows significant potential as an assistive tool for clinical diagnosis in prostate cancer.
• Further refinement is needed to improve the specificity, particularly for physiological uptake in glands.