Keywords:
Artificial Intelligence, Computer applications, Breast, CAD, CT, Neural networks, Computer Applications-Detection, diagnosis, Screening
Authors:
X. Tian1, R. Zhang1, C. Xia1, S. Liang1, L. Yang2, Y. Zhu2; 1Beijing/CN, 2Guangzhou/CN
DOI:
10.26044/ecr2019/C-1202
Methods and materials
Our model predicts the stages based on the ROI (region of interests) from CT images,
and a modified DenseNet [3] is designed to extract the 3D features of ROI for classification.
Data Description:
The thymoma dataset has total number of 184.
The training set consists of 147 chest CT sequences (80%),
and the test dataset has 37 chest CT sequences (20%).
The gold standard was acquired from institutional database for medical records which identified patients with histologically confirmed thymic epithelial tumor who underwent surgical resection,
where 15 cases were confirmed by biopsy,
and 169 cases were confirmed by postoperative pathology.
Data Augmentation:
Data augmentation is adopted to avoid overfitting.
There are two types of data augmentations.
First,
random cropping was adopted to make variants of the training data.
Second,
random window width and center were applied in our training samples.
Data augmentation is essential to overcome overfitting because our datasets is small.
Also,
by using data augmentation,
model performance can gain extra value from the augmented data.
3D-DenseNet Model:
We use a novel model architecture to analyze the thymoma data.
The proposed model was derived from DenseNet [3].
DenseNet is composed with several Dense Blocks,
as shown in Fig.
1.
In each Dense Block,
each layer connects every other layers in a feed-forward fashion (Fig.
2).
This feature makes DenseNet the ideal model to have more deep layers.
DenseNet was chosen as the base model for several reasons.
First,
deep model has better performance in fitting data.
Second,
the Dense Block components in it are beneficial to overcome overfitting problem.
The proposed model has 169 layers,
and this study has two contributions to the original DenseNet.
First,
we changed the common DenseNet from two-dimensional form to three-dimensional form by modifiying the kernels to 3D version.
In this way,
the model was tailored for CT volume data.
Then,
we adopted transfer learning for this 3D-DenseNet model,
using parameters from ImageNet [4].
To achieve this,
the pretrained parameters from 2D DenseNet were extended by copying them to another dimension.
The resulsts has shown that transfer learning in this way can boost the training speed and make training process more stable.
Implementation:
The proposed model was trained under the deep learning framework MXNet.
We used binary cross-entropy as the loss function,
and Adam as the optimizer to speed up training.
To increase training performance,
cosine annealing was adopted to schedule the learning rate for 1200 steps,
using learning rate ranging from 1e-5 to 1e-7.
We set mini-batch with size of 16.
The training was conducted using 4 NVIDIA GeForce GTX 1080 GPUs (NVIDIA,
Beijing,
China).