Application of Deep Learning to Predict Standardized Uptake Value Ratio and Amyloid Status on 18F-Florbetapir PET Using ADNI Data

Article Figures & Data

Download figure
Open in new tab
Download powerpoint
Fig 1.
A, The input to ResNet consists of 3 or more input channels. In the case of 1 section prediction, the section is copied to all 3 color channels. If 3 slices are used as input, each color channel has an individual section. The input layer can be modified to include more slices as well. The convultional neural network can be used to predict amyloid status directly or to measure SUVR (regression). B, Histogram of all SUVR values from the cases included in this study (n = 2582).
Download figure
Open in new tab
Download powerpoint
Fig 2.
ROC curves test results. A, shows performance for 1-section and 3-section input data for binary classification. B, displays performance for 1- and 3-section classification via regression. All configurations use pretrained ImageNet weights.
Download figure
Open in new tab
Download powerpoint
Fig 3.
Classification performance as a function of number of input slices. All results reflect the average of 5 seeded runs.
Download figure
Open in new tab
Download powerpoint
Fig 4.
Examples of PET scans used for single section prediction. The top number represents the prediction of the network while the bottom number is the ground truth manual SUVR measurement from the ADNI data base.

Table 1:

Various test metrics for binary classification. Performance reflects mean of 5 separate seeded runs

Binary Classification	Accuracy	Sensitivity	Specificity	PPV	NPV	ROC AUC
1 section, random initialization	92.83% (1.16%)	89.76% (2.11%)	95.87% (0.43%)	95.61% (0.68%)	90.34% (1.69%)	0.9735 (0.0095)
1 section, pretrained	93.41% (1.13%)	91.07% (2.39%)	95.69% (0.72%)	95.55% (0.46%)	91.48% (2.00%)	0.9815 (0.0059)
3 sections, random initialization	92.40% (0.86%)	90.24% (1.05%)	94.56% (0.92%)	94.32% (1.17%)	90.59% (0.82%)	0.9782 (0.0072)
3 sections, pretrained	93.88% (0.78%)	91.52% (1.40%)	96.23% (0.94%)	96.14% (0.64%)	91.80% (1.71%)	0.9850 (0.0044)
9 sections, random initialization	93.14% (0.95%)	90.79% (2.19%)	95.39% (1.77%)	95.26% (1.81%)	91.28% (1.12%)	0.9821 (0.0064)
9 sections, pretrained	93.14% (0.69%)	91.75% (1.77%)	94.43% (1.85%)	94.43% (1.62%)	92.01% (1.18%)	0.9843 (0.0038)
27 sections, random initialization	93.84% (1.04%)	92.23% (1.88%)	95.40% (1.54%)	95.31% (1.48%)	92.49% (1.46%)	0.9831 (0.0062)
27 sections, pretrained	93.45% (0.64%)	90.67% (1.58%)	96.17% (1.19%)	96.01% (1.15%)	91.17% (0.86%)	0.9858 (0.0041)

Table 2:

Linear mixed-effects model analysis of different methods for classifying amyloid PET imaging

Factor	Accuracy		Sensitivity		Specificity
Factor	Odds Ratio	P Value	Odds Ratio	P Value	Odds Ratio	P Value
Type (ResNet-50 vs. -152)	0.96 (0.89–1.03)	.258	0.916 (0.82–1.02)	.112	1.00 (0.89-1.11)	.944
Method (binary classification vs. regression first)	0.79 (0.70–0.89)	<.001	0.54 (0.45–0.63)	<.001	1.24 (1.03–1.50)	.024
Initialization (random vs. pretrained)	0.33 (0.30–0.37)	<.001	0.45 (0.38–0.53)	<.001	0.22 (0.19–0.26)	<.001
Slices (1 vs. 3)	1.17 (1.08–1.26)	<.001	1.15 (1.03-1.28)	.012	1.22 (1.10–1.36)	<.001

Table 3:

Comparison of prediction performance for 100 randomly selected test set samples

Clinical Evaluation	ResNet-50	Reader 1	Reader 2	Reader 3	All Readers
Accuracy	96.00%	90.00%	86.00%	89.00%	90.00%
Sensitivity	95.83%	85.42%	77.08%	87.50%	85.42%
Specificity	96.15%	94.23%	94.23%	90.38%	94.23%
PPV	95.83%	93.18%	92.50%	89.36%	93.18%
NPV	96.15%	87.50%	81.67%	88.68%	87.50%
Time	0:03 min	8:00 min	9:30 min	6:58 min	24:28 min