Computer-aided bone scan lesion area quantitation: Inter-reader measurement variability.

Abstract

e16019 Background: Bone scan is essential for response assessment in subjects with metastatic prostate cancer. However, visual assessment is subjective and prone to inter-reader variability. A commercial computer-aided (CAD) system for bone scan lesion area (BSLA) calculation has been developed (MedQIA, LLC). BSLA provides an objective and sensitive measure for response assessment. New VEGF treatments showing promising results in reducing lesion burden on bone scan present an urgent need for validated systems capable of accurately assessing treatment response. Measurement of inter reader variability is fundamental to such a validation. In this work we measure inter-reader variability in quantitation of BSLA change using the CAD system. Methods: In a multi-center clinical trial of cabozantinib (Exelixis, Inc), bone scans were acquired according to a standardized imaging protocol. Scans at baseline and Week 12 were available for 113 subjects for this analysis. Automated CAD lesion detection was performed following image intensity normalization to increase consistency between subject time points. Two nuclear medicine physicians independently edited and approved the CAD result as needed, primarily to remove false positives due to degenerative joint disease. The system then calculated the percent change in BSLA at Week 12 from baseline (BSLAPCT). The BLSAPCT values from the two readers were compared to assess inter-reader variability. Additional analytic validation of the BSLA measure was performed in which the readers provided a categorical assessment of responder vs non-responder, using the CAD-normalized images, and the classification performance of BSLAPCT was assessed using ROC analysis. Results: The median (IQR) difference in BSLAPCT between the two readers was -1.8 (20.0) percentage points. The absolute difference was 9.4 (23.9) percentage points. The BSLAPCT response classification had an ROC area under curve (AUC) of 0.96 for both readers. Conclusions: Computer-aided quantitative assessment of change using bone scan lesion area demonstrated low variability between two readers. This is an important step toward the validation of an objective, quantitative, and sensitive response assessment technique.