Noninvasive Prediction of Occult Peritoneal Metastasis in Gastric Cancer Using Deep Learning

Key Points Question Can occult peritoneal metastasis be accurately assessed before surgery and without any invasive intervention? Findings In this cohort study of 1978 patients, a deep neural network, the Peritoneal Metastasis Network, was developed for predicting occult peritoneal metastasis in gastric cancer based on preoperative computed tomography images. The model had excellent discrimination in external validation and substantially outperformed clinical factors. Meaning The proposed deep learning model may be useful in preoperative treatment decision-making for avoiding unnecessary surgery and complications in certain patients.


CT Acquisition and Image Processing
All patients underwent contrast-enhanced abdominal CT using the multidetector row CT (MDCT) systems (GE Lightspeed 16, GE Healthcare Milwaukee, WI; 64-section LightSpeed VCT, GE Medical Systems, Milwaukee, WI; USA). Following intravenous contrast administration, arterial and portal venous-phase contrast-enhanced CT scans were performed after delays of 28 s and 60 s, respectively. Iodinated contrast material in the amount of 90 -100 ml (Ultravist 370,Bayer Schering Pharma,Berlin,Germany) was injected at a rate of 3.0 or 3.5 ml/s with a pump injector (Ulrich CT Plus 150, Ulrich Medical, Ulm, Germany). The CT acquisition protocols were as follows: 120 kV; 150-190 mAs; 0.5-or 0.4-second rotation time. Contrast-enhanced CT was reconstructed with a field of view, 350×350 mm; data matrix, 512×512; in-plane spatial resolution 0.607-0.751 mm; axial slice thickness 5.0 mm for 98% patients with a range of 1.25-7.5 mm.
We analyzed the portal venous-phase CT images because of well differentiation between the tumor tissue and adjacent normal bowel wall. The relatively coarse and heterogeneous resolution in z-axis compared with in-plane resolution would not allow a meaningful and reliable 3D analysis of the image. Therefore, we focused on the most representative 2D slice, i.e., largest tumor section in the axial plane. Two radiologists C.C. and Q.Y. (with 11 and 10 years of clinical experience in abdominal CT interpretation, respectively) manually delineated the primary tumor on the CT images by using the ITK-SNAP (http://www.itksnap.org) 1, 2 .

Network architecture
The proposed DCCN-LSC deep learning model ( Figure 1A) consists of a convolutional layer, two dense blocks each followed by a transition layer, and a final dense block followed by a pooling and a linear layer. The dense blocks ( Figure 1B) use short dense connectivity between sequences of convolution, batch normalization, and rectified linear units (ReLU). The transition layers, formed by a convolutional and a pooling layer, are used to reduce the dimension of the feature maps between adjacent dense blocks. The final pooling and linear layers are used to reduce the output dimension to PM prediction. All convolutional operators use a stride of 2 and the kernel size of 3. After the CT image input, we add a convolutional layer with the 2×2×2 stride.
The convolutional layer is followed by four dense blocks, which use dense connectivity formed by the output feature from all the prior layers: where [ 0 , 1 , … , −1 ] is the tensor that concatenating the feature maps from all previous layers.
is a non-linear transformation function of three sequential processes: convolution, batch normalization, and rectified linear units. We use the shortcut connection to enable the dense layer can receive the feature maps from all the previous dense layers. We set a transition layer to reduce the dimension of the feature maps between the adjacent dense blocks. The transition layer is formed by a convolutional and pooling layer. To make a regression to the occult PM prediction, we add a pooling and a linear layer to the last dense block for reducing the dimension of the feature map.
Different from traditional dense-net that only with the short connection inside the dense blocks, DCCN-LSC introduces a long connection that enables the model to extract the multi-level feature of the tumor. The multi-level feature maps are incorporated into the final fully connection layer for accurate occult PM prediction.

Data augmentation
As Gaussian filtering and the noise selects a standard deviation of the Gaussian distribution randomly in the range specified by the minimum and the maximum.

Implementation detail
The loss function of the occult PM prediction is binary cross-entropy. To minimize the loss function, we use Adam algorithm to obtain the optimal parameters. The learning rate is set at 10 -2 initially and then gradually decreased slowly decrease it to 10 -6 . The DCNN-LSC model was trained for 100 epochs with a batch size of 16. We train the data using Matlab on 4 NVIDIA GeForce GTX 1080 Ti GPUs, an Intel Xeon(R) CPU E5-1650 v4 @ 3.60GHz × 12, and 64 GB of internal memory.