Liver cancer is a leading cause of cancer deaths worldwide, accounting for more than 600,000 deaths each year. According to the American Cancer Society’s estimates for primary liver cancer and intrahepatic bile duct cancer, in 2018 in the United States are expected:
- About 42,220 new cases to be diagnosed
- About 30,200 people to die of these cancers
Skychain strives to help doctors detect the liver cancer as early as possible and we already have some good results on our neural network to share. Skychain developer Viktoriya Akhmadieva gave us some information about her ANN for liver cancer detection by CT scans.
The neural network’s task is to detect liver cancer by analysing computed tomography images. For this purpose, the main task is divided into three stages: segmentation of a liver on CT images, detection of tumours and their segmentation. There is one neural network for each subtask created. They are executed sequentially, working with the results obtained from the previous neural network.
There are 4 stages of liver cancer and each of them is characterized by the presence of metastases and tumours of different size and number. The ANN’s objective is to determine whether the liver is healthy or in the opposite case to segment a tumour.
According to the results of liver segmentation, the detector locates the tumour in the image. For this, the neural network based on the fully convolutional network (FCN) ResNet50 with the binary cross entropy as loss function is used.
Then, in order to segment the tumour, the next ANN trains with backpropagation algorithm. This neural network is based on Vgg16 with the sigmoid activation function and MomentumOptimizer.
This is the loss function graph of the Skychain Neural Network for Liver Cancer Detection below. The graph shows the reduction in the number of prediction errors during the training period of the neural network.