Training method and device of neural network for medical image processing, and medical image processing method and device

A training device of a neural network for medical image processing, comprising:

• at least one hardware processor;
• and
• a memory having computer program instructions stored thereon that, when executed by the at least one hardware processor, direct the at least one hardware processor to:
• perform a pre-processing process on an original image to obtain a pre- processed image;
• perform a data-augmenting process on the pre-processed image to obtain an augmented image which retains a pathological feature, the augmented image comprising at least one image with a first resolution, at least one image with a second resolution and at least one image with a third resolution, and the second resolution and the third resolution being higher than the first resolution and being not equal to each other;
• and
• train the neural network by inputting the at least one image with the first resolution, a center-cropping image from the at least one image with the second resolution and a center-cropping image from the at least one image with the third resolution respectively into three different input terminals of a single neural network as training samples at the same time,
• wherein resolutions of the center-cropping image from the at least one image with the second resolution, the center-cropping image from the at least one image with the third resolution and the at least one image with the first resolution are the same,
• wherein the at least one image with the first resolution, the center-cropping image from the at least one image with the second resolution and the center-cropping image from the at least one image with the third resolution are derived from the same original image.

A medical image processing method based on a neural network, comprising:

• obtaining a medical image;
• processing the medical image by inputting the medical image into the neural network trained by:
• performing a pre-processing process on the medical image to obtain a pre-processed image;
• performing a data-augmenting process on the pre-processed image to obtain an augmented image which retains a pathological feature, the augmented image comprising at least one image with a first resolution, at least one image with a second resolution and at least one image with a third resolution, and the second resolution and the third resolution being higher than the first resolution and being not equal to each other;
• and
• training the neural network by inputting the image with first resolution, a center-cropping image from the image with second resolution and a center-cropping image from the at least one image with the third resolution respectively into three different input terminals of a single neural network as training samples at the same time;
• and
• outputting a processed result,
• wherein resolutions of the center-cropping image from the at least one image with the second resolution, the center-cropping image from the at least one image with the third resolution and the at least one image with the first resolution are the same,
• wherein the at least one image with the first resolution, the center-cropping image from the at least one image with the second resolution and the center-cropping image from the at least one image with the third resolution are derived from the same original image.

A medical image processing device based on a neural network, comprising:

• at least one hardware processor;
• and
• a memory having computer program instructions stored thereon that, when executed by he at least one hardware processor, direct the at least one hardware processor to:
• obtain a medical image;
• process the medical image by inputting the medical image into the neural network trained by:
• performing a pre-processing process on the medical image to obtain a pre- processed image;
• performing a data-augmenting process on the pre-processed image to obtain an augmented image which retains a pathological feature, the augmented image comprising at least one image with first resolution, at least one image with second resolution and at least one image with a third resolution, and the second resolution and the third resolution being higher than the first resolution and being not equal to each other;
• and
• training the neural network by inputting the image with first resolution, a center-cropping image from the image with second resolution and a center-cropping image from the at least one image with the third resolution respectively into three different input terminals of a single neural network as training samples at the same time;
• and
• output a processed result,
• wherein resolutions of the center-cropping image from the at least one image with the second resolution, the center-cropping image from the at least one image with the third resolution and the at least one image with the first resolution are the same,
• wherein the at least one image with the first resolution, the center-cropping image from the at least one image with the second resolution and the center-cropping image from the at least one image with the third resolution are derived from the same original image.