System and method for segmenting three- dimensional images of pancreases and tumors

A method for segmenting three-dimensional (3D) images of pancreases and tumors, comprising:

• acquiring 3D images of the pancreas and the tumor and preprocessing same with a soft tissue window to control an intensity value of the image within a set range;
• cropping all images into block-shaped regions of a set size and feeding same into a convolutional neural network model, and when training a network, dynamically adjusting a learning weight between the pancreas and the tumor under the guidance of temperature to learn features of the pancreas and the tumor;
• after performing data preprocessing on the 3D images of the pancreas and labels, using a trained network model for online testing and evaluating, and outputting the segmentation result;
• and
• using a balanced temperature loss (BTL) to dynamically adjust the learning weight according to the temperature, resulting in a good performance of segmentation of both the tumor and the pancreas;
• under an influence of a deep supervision mechanism, BTL being expressed as:
• B
• ⁢
• T
• ⁢
• L
• =
• ∑
• i
• =
• 1
• 4
• ω
• i
• ∑
• i
• =
• 1
• 4
• ω
• i
• ⁢
• L
• i
• ,
• ω
• i
• =
• 1
• 2
• i
• ;
• where, Li is the loss of each layer of a decoder, i=1,2,3,4;
• the network being trained in combination with a binary cross-entropy loss function and the soft Dice loss as a loss function of an NAD skeleton to constrain learning, and temperature being introduced in the soft Dice loss to dynamically adjust the learning weight between the tumor and the pancreas to balance feature selection;
• and
• the pancreas being given a large weight under the guidance of temperature at an early stage of network learning, and then a tumor region being dynamically given a larger weight as the network learning continues, so that the features of the pancreas and the tumor can be dynamically learned in a process of network learning.