Kidney and kidney tumor segmentation method based on deep neural network

Abstract

The invention relates to a kidney and kidney tumor segmentation method based on a deep neural network, and the method comprises the following steps: constructing a CT medical image data obtaininga CT medical image data set of a kidney cancer patient, and selecting an image containing a kidney and kidney tumor position region manually marked by a doctor in the data set as a training set; training a roughly segmented U-Net + + neural network; carrying out image enhancement; training a neural network of the improved Dense-Unet; sending the image into an improved Dense-Unet neural network tocarry out final segmentation training; in the training process, using an Adam optimizer to modify a learning rate gradient descent method into a Cosine function descent method, enabling a loss function to utilize a loss function combining Dice Loss with CrossEntropyLoss, wherein a performance evaluation index of the neural network is a Dice index.

Description

technical field [0001] The invention belongs to the field of image segmentation, and relates to a multi-category segmentation method for three-dimensional CT kidney medical images by using a neural network. Background technique [0002] Kidney and kidney tumor segmentation is a technical means to separate the kidney organ area and kidney tumor area from other healthy tissues in the three-dimensional medical images taken by patients with kidney cancer. [1] It pointed out that in 2018, more than 400,000 people worldwide were diagnosed with kidney cancer, and more than 175,000 patients died of kidney cancer. Using enhanced CT images to diagnose kidney cancer is a key and necessary step in medicine. Using image segmentation to automatically label patients' CT images can assist doctors in the diagnosis of tumor lesions in the kidney area, reducing the misdiagnosis rate and saving money. The time that doctors manually mark the tumor area can win treatment time for patients. [0...

AI Technical Summary

Problems solved by technology

In the task of segmenting kidneys and kidney tumors, there is a problem with CT images: most of the data slices do not contain kidneys or kidney tumors, so if all the slices are sent to the neural network for training, there will be a lot of meaningless training , there will be a lot of redundant information

In the segmentation process of medical images, if the convolution kernel only uses ordinary convolution, it will lead to the loss of internal data structure and the loss of spatial level information

Images