Table 4. Candidate CNN layers from the coarse models showing superior performance with the RSNA CXR test set.

The performance metrics achieved through feature extraction from these intermediate layers of the coarse models helped to achieve superior performance toward classifying normal and abnormal CXRs using the RSNA CXR test set. The naming conventions for the models layers follow that available from the Keras DL library.

Model	Feature extraction layer
Custom CNN	Conv3
VGG-16	Block5-conv3
VGG-19	Block4-pool
Inception-V3	Mixed3
Xception	Add-7
DenseNet-121	Pool3-conv
MobileNet	Conv-pw-6-relu
NASNet-mobile	Activation-129