Diagnostic performance of deep learning models for detecting bone metastasis on whole-body bone scan in prostate cancer

Abstract

Purpose: We evaluated the performance of deep learning classifiers for bone scans of prostate cancer patients.

Methods: A total of 9113 consecutive bone scans (5342 prostate cancer patients) were initially evaluated. Bone scans were labeled as positive/negative for bone metastasis using clinical reports and image review for ground truth diagnosis. Two different 2D convolutional neural network (CNN) architectures were proposed: (1) whole body-based (WB) and (2) tandem architectures integrating whole body and local patches, here named as "global-local unified emphasis" (GLUE). Both models were trained using abundant (72%:8%:20% for training:validation:test sets) and limited training data (10%:40%:50%). The allocation of test sets was rotated across all images: therefore, fivefold and twofold cross-validation test results were available for abundant and limited settings, respectively.

Results: A total of 2991 positive and 6142 negative bone scans were used as input. For the abundant training setting, the receiver operating characteristics curves of both the GLUE and WB models indicated excellent diagnostic ability in terms of the area under the curve (GLUE: 0.936-0.955, WB: 0.933-0.957, P > 0.05 in four of the fivefold tests). The overall accuracies of the GLUE and WB models were 0.900 and 0.889, respectively. With the limited training setting, the GLUE models showed significantly higher AUCs than the WB models (0.894-0.908 vs. 0.870-0.877, P < 0.0001).

Conclusion: Our 2D-CNN models accurately classified bone scans of prostate cancer patients. While both showed excellent performance with the abundant dataset, the GLUE model showed higher performance than the WB model in the limited data setting.