Invented by Yingying Li, Lvwei WANG, BOE Technology Group Co Ltd
The BOE Technology Group Co Ltd invention works as follows
The present disclosure includes a method for training a neural net for medical imaging processing, as well as a method for processing medical images based on the neural network. The training method comprises performing a preprocessing on an original image in order to obtain a processed image, performing an image-augmenting on the processed image to produce an augmented picture retaining a feature pathological, the augmented picture including at least an image with a first resolution (and at least an image with a second resolution higher than the first), and training the neural system by selecting an image of first resolution from the image of second resolution.Background for Medical image processing method and devices
Early detection and accurate diagnoses of skin diseases, particularly melanoma, are crucial. Dermoscopy is also called skin surface microscope. It’s a non-invasive technique that uses microscopic images to observe microstructures and skin pigments. The dermoscopy, also known as skin surface microscope, is a non-invasive microscopic image analysis technique for observing microstructures and pigments under the skin surface of living body.
It should be noted that information in the previous background section may not constitute prior knowledge that is already well known by those with ordinary skill in this art.
The present disclosure relates to a method and device for medical image analysis, as well as a method and apparatus for processing medical images based on neural networks.
The present disclosure provides a method for training a neural net to process medical images. This includes performing a preprocessing on an original image in order to obtain a “pre-processed” image, performing an “augmenting” process on the preprocessed picture to obtain an “augmented” image that retains the pathological feature.
In an arrangement of the current disclosure, performing a preprocessing process on an image original includes performing a pixel-normalizing process and a colour constancy process on the image original.
In an arrangement of the current disclosure, performing a data enhancement process on the image pre-processed includes at least one process to enhance the image pre-processed: cropping, rotating, upside-down, inverting, and horizontal inverting processes.
In an arrangement of the current disclosure, performing a data enhancement process on a pre-processed picture includes also performing an image warping process on that pre-processed picture.
The arrangement described in the present disclosure includes an original image that contains a dermoscopy, a cropping process on the preprocessed images to maintain a boundary around a lesion, and an image-warping on the preprocessed images to retain the symmetrical characteristics of the lesion.
In an arrangement of this disclosure, the part cropping image from the second resolution image includes a centre-cropping, and the same resolution is used for the center-cropping as the first resolution image.
The neural network is trained using the images with the first and second resolutions, as well as a part-cropping from the second resolution.
In an arrangement of this disclosure, the resolutions of the images of the image of second resolution, of the image of third resolution, and of the first resolution image are all the same.
The present disclosure provides a device for training a neural net to process medical images, which includes: a preprocessing part configured to perform preprocessing on an original image in order to obtain a processed image; an augmenting component configured to perform data-augmenting on the preprocessed picture to obtain an enhanced image that retains the pathological feature. This augmented image comprises at least two images, one with a first resolution and one with a second resolution. The second resolution is higher than the resolution of the pre
The arrangement in the present disclosure includes two augmenting parts. The first is configured to perform one or more of the following processes in order to enhance the preprocessed images: cropping, rotating, upside-down and horizontal inverting, while the second is configured for an image warping process.
The neural network is trained by introducing the image of first resolution as well as the part-cropping from the image of second resolution.
In an arrangement of this disclosure, the resolutions of the images of the image of second resolution, of the image of third resolution, and of the first resolution image are all the same.
The present disclosure describes a device that can be used to train a neural network to process medical images. It includes a processor and a memory with computer program instructions. When executed by the processor these instructions cause the processor to execute one or more blocks of the training method.
The present disclosure provides a method of processing medical images based on neural networks, which includes: obtaining a digital medical image; processing it by putting the image into the trained neural network by using the training method above; and displaying the processed result.
In an arrangement of the current disclosure, processing the medical images includes classification of the medical images based on a feature pathological of a lesion.
The present disclosure describes a medical image-processing device that is based on neural networks. It includes: a processor and a storage memory with computer program instructions. When executed by the processor these instructions cause the processor perform one or several blocks of the medical image-processing method.
The computer program instructions, when executed by the processor, cause the processor perform the above method of medical image processing to classify a medical image based upon a pathological characteristic of a lesion in the medical picture.
According to an arrangement in the present disclosure, an electronic apparatus is provided for medical image-processing based on neural networks, which includes: an image-obtaining part configured to obtain a clinical image; an image-processing part configured to process the image using the neural network trained by this training method; and an output part configured to output a result of image processing.