Table 2. Datasets used for phase II deep convolutional neural network training and validation.
Dataset | Development phase | Total number of radiographs used | Number with suspected TB (%) | Number without suspected TB (%) |
---|---|---|---|---|
NIH CXR | Training/validation | 100,622 | 44,521 (44.2)+ | 56,101 (55.8)+ |
Montgomery County, USA | Testing | 138 | 58 (42.0) | 80 (58.0) |
Shenzhen, China | Testing | 662 | 336 (50.8) | 326 (49.2) |
Johns Hopkins Hospital, Baltimore, MD, USA | Testing | 100 | 35 (35.0) | 65 (65.0) |
+, ground truth determined by phase I deep learning system (DLS). CXR, chest radiograph; TB, tuberculosis.