A Weisman¹*, M Kieler¹ , S Perlman¹ , R Jeraj^1,2 , M Hutchings³ , L Kostakoglu⁴ , T Bradshaw¹ , (1) University of Wisconsin-Madison, Madison, WI, (2) Faculty of Mathematics and Physics, Ljubljana, Slovenia, (3) Rigshospitalet, Copenhagen, Denmark, (4) Mount Sinai Medical Center, New York, NY

• 
  A Weisman

(Wednesday, 7/17/2019) 10:15 AM - 12:15 PM

Room: 304ABC

Purpose: Quantification of ¹�F-FDG PET/CT images can improve prognostication and treatment response assessment for lymphoma patients. However, full quantification is difficult because high disease burden is often present. Here, a fully-automated method is developed and assessed for its ability to quantify lymphoma in FDG PET/CT images.

Methods: Baseline disease on FDG PET/CT images of 90 lymphoma patients was segmented by a nuclear medicine physician. Images were cropped to consider only disease above the diaphragm. Cascaded patch-based 3D dual-resolution pathway convolutional neural networks (CNNs) were trained on the PET/CT images using 5-fold cross-validation. The first stage was an ensemble of CNNs trained to detect disease. Its output was fed to a second stage, where disease boundaries were segmented for quantification. Patient-level SUVmax, SUVtotal, and PET volume were compared between automated and physician-based quantification using Pearson’s correlation coefficient. Results were compared to the inter-physician agreement of 20 patients that were segmented by a second nuclear medicine physician.

Results: Across all 90 patients, the physician identified 776 lesions (range:0-47/patient) above the diaphragm, with median SUVmax of 7.7g/mL (range:1.1-34.8) and PET volume of 1.6cm3 (range:0.008-719.1). The automated method achieved a detection sensitivity of 87% at 3 false positives (FPs)/patient, and median lesion-level Dice of 0.64 (interquartile range:0.43-0.76) in true positive lesions. Strong to excellent agreement was found for all metrics: R=0.99 for SUVmax, R=0.90 for SUVtotal, and R=0.85 for volume. In the 20 patients segmented by both physicians (N=154 lesions), automation achieved performance comparable to physician agreement, with detection sensitivity of 93% (inter-physician:97%) at 3.4 FPs/patient, and median lesion-level Dice of 0.63 (inter-physician:0.64). Correlation coefficients were 0.99 for SUVmax (inter-physician:1.0), 0.76 for SUVtotal (inter-physician:0.90), and 0.72 for volume (inter-physician:0.60).

Conclusion: A fully-automated, cascaded CNN-based method achieved FDG PET/CT quantification performance above the diaphragm comparable to inter-physician agreement in a heterogeneous lymphoma patient population.

Funding Support, Disclosures, and Conflict of Interest: This research was supported by GE Healthcare.

Keywords

CAD, PET, Quantitative Imaging

Taxonomy

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