Fig 1.
CT examples with lung nodules in different categories.
They are benign (left), primary malignant (middle) and metastatic malignant (right), alone with 3 different view areas including 40×40, 50×50 and 60×60.
Fig 2.
Two kinds of the multi-view strategy.
The one-view-one-network strategy (top) employs a separate network for images from each view (cropping size), while the multi-view-one-network strategy (bottom) uses one network for all views.
Fig 3.
The architectures of the 3D single view CNN (SV-CNN) (top) and 3D MV-CNN (bottom).
Table 1.
The configuration of the network with chain architecture.
Fig 4.
The overall architecture of the 3D multi-view Inception network (left). This architecture applies to Inception1 and Inception2. The output sizes of each layer for the binary classification (middle) and the ternary classification (right) are different, which are also shown in this figure. Note that the output of each layer is a 5D tensor as described in the previous section. The details of the Inception module are shown in Fig 5.
Fig 5.
The detail of the Inception module in Fig 4: 3D Inception1 (left) and 3D Inception2 (right).
Fig 6.
The overall architecture of the Inception-ResNet network.
Fig 7.
The process of generating 3D volume data.
Table 2.
The number of patients and lesions.
Fig 8.
The error rate of the 3D SV-CNN (top) and 3D MV-CNN (bottom) with chain architecture for the binary classification.
From left to right, there are the error rates of Softmax, CNN1, CNN2 and CNN3, respectively.
Fig 9.
The error rate of the 3D SV-CNN (top) and 3D MV-CNN (bottom) with chain architecture for the ternary classification.
From left to right, there are the error rates of Softmax, CNN1, CNN2 and CNN3, respectively.
Table 3.
The result of the binary classification for the networks with chain architecture.
Table 4.
The result of the ternary classification for the networks with chain architecture.
Table 5.
The validation error rate of the classification for the 3D multi-view networks with DAG architecture.
Table 6.
Some classification results on LIDC-IDRI dataset.
Fig 10.
The ROC curves of MV-softmax, 3D MV-CNN1 and 3D MV-inception1.
Fig 11.
The number of parameters vs. training time and validation error rate for the binary classification (left) and the ternary classification (right).
The area of the circles represents the corresponding validation error rate.
Fig 12.
The result of grid search for the binary (left) and ternary (right) classification.
Table 7.
Compare the multi-view-one-network strategy with the one-view-one-network strategy.