Click here to


Are you sure ?

Yes, do it No, cancel

Comparison of Automated Methods for Detection and Prognostication in Metastatic Prostate Cancer Using 18F-NaF PET/CT Images

B Schott*1, A J Weisman1, T G Perk1, A R Roth1, S Yip2, G Liu1, R Jeraj13, (1) University of Wisconsin, Madison, WI, (2) AIQ Solutions, Madison, WI, (3) University of Ljubljana, Slovenia


(Wednesday, 7/15/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose: Quantitative ¹8F-NaF PET/CT image metrics are prognostic in metastatic prostate cancer (mPC) but difficult to implement due to widespread disease and high-uptake benign lesions. Here, we automatically detect and quantify malignant bone disease on ¹8F-NaF PET/CT images of mPC patients using multiple image analysis techniques and compare the prognostic power of each technique.

Methods: Nuclear medicine physicians contoured malignant bone lesions on ¹8F-NaF PET/CT images of 37 mPC patients, identifying 1070 lesions. PET/CT images were used as inputs for three established disease detection
methods: SUV>10; SUV>15; Statistically-Optimized Regional Thresholding with random forest classification (SORT+RF); and a newly implemented 3D convolutional neural network (DeepMedic) was tested. Median detection sensitivity and number of false positives (FPs) per patient were investigated. SUVmax, SUVmean, SUVtotal, and number of lesions (NL) were extracted from patients. The prognostic power of each method was assessed using concordance (C-index) and hazard ratios (HR) of independent variables from multivariate Cox proportional-hazards regression models of progression-free survival.

Results: Detection analysis showed a median sensitivity of 91% at 31 FPs/patient and 74% at 4 FPs/patient for SUV>10 and SUV>15, respectively. At similar FPs/patient, DeepMedic had significantly better sensitivity of 100% and 79%, respectively (p<0.002, Wilcoxon paired test). At 5 FPs/patient, DeepMedic had significantly higher sensitivity than SORT+RF (79% vs. 75%, p = 0.004). Method-specific prognostic analysis yielded the following multivariate variables: SUV>10 (C-index=0.71) with SUVmean (HR=2.13, p<0.001); SUV>15 (C-index=0.70) with SUVmean (HR=1.88, p=0.006) and SUVtotal (HR=1.37, p=0.08); SORT+RF (C-index=0.77) with SUVmean (HR=2.35, p<0.001) and NL (HR=2.26, p=0.009); and DeepMedic at 5 FPs/patient (C-index=0.71) with SUVmean (HR=1.67, p=0.06) and NL (HR=2.42, p<0.001).

Conclusion: DeepMedic-based disease contours showed superior malignant detection compared to all previously established methods. DeepMedic-based prognostic performance was comparable to fixed thresholding methods but was outperformed by SORT+RF. This motivates further work to optimize disease detection for prognostication.

Funding Support, Disclosures, and Conflict of Interest: Data was acquired through funding from the Prostate Cancer Foundation. This work was additionally supported by the University of Wisconsin Carbone Cancer Center Support Grant P30 CA014520. Author SY is CIO, GL is CMO, and RJ is CSO of AIQ Solutions. Authors GL and RJ are co-founders of AIQ Solutions.


Quantitative Imaging, Image Analysis, PET


IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email