Method and apparatus for analyzing medical image
The present disclosure discloses a method and apparatus for analyzing medical image. A specific embodiment of the method includes: acquiring medical image data; generating multi-scale decision sample data based on the medical image data; inputting the multi-scale decision sample data into a deep neural network model to obtain an auxiliary diagnosis data of the medical image, the deep neural network model being trained according to a consistency principle between multi-scale training sample data and an output result of the deep neural network model. In the embodiment, a multi-scale training sample is used to accelerate the training process of the deep neural network model, thus the auxiliary diagnosis decision process can be accelerated, while the accuracy of the trained deep neural network model of the embodiment is improved according to a consistency principle of data between different scales and output results, thereby improving the accuracy of auxiliary diagnosis decision.
1. A method for analyzing a medical image, the method comprising:
acquiring medical image data;
generating multi-scale decision sample data based on the medical image data; and
inputting the multi-scale decision sample data into a deep neural network model to obtain auxiliary diagnosis data of the medical image, wherein the deep neural network model is trained using multi-scale training sample data according to a consistency principle between the multi-scale training sample data and an output result, wherein the deep neural network model is trained by:
acquiring full-field pathological image data;
generating the multi-scale training sample data based on the full-field pathological image data; and
training the deep neural network model according to a consistency principle between the multi-scale training sample data and the output result of the deep neural network model using the multi-scale training sample data,
wherein the generating the multi-scale training sample data based on the full-field pathological image data comprises:
removing a non-effective pathological tissue area in the full-field pathological image data to obtain an effective pathological tissue area; and
generating the multi-scale training sample data based on the effective pathological tissue area, and
wherein the method is performed by at least one processor.
2. The method according to claim 1 , wherein the generating the multi-scale training sample data based on the full-field pathological image data comprises:
determining a positional correspondence relationship between the multi-scale training sample data using a multi-scale alignment algorithm.
3. The method according to claim 1 , wherein the training the deep neural network model using the multi-scale training sample data comprises:
training the deep neural network model respectively using training sample data of different scales in the multi-scale training sample data; or
training the deep neural network model jointly using the training sample data of different scales in the multi-scale training sample data.
4. The method according to claim 1 , wherein the training the deep neural network model using the multi-scale training sample data comprises:
establishing a connection between training sample data of different scales in the multi-scale training sample data using topology or dynamic coding; and
training the deep neural network model based on the connection between the training sample data of different scales.
5. The method according to claim 1 , wherein the training the deep neural network model using the multi-scale training sample data comprises at least one of:
selecting sample data with a score lower than a predetermined threshold from the multi-scale training sample data for back propagation, to update parameters of the deep neural network model; or
selecting sample data expired in at least one of training or detection from the multi-scale training sample data for hard example mining, to update parameters of the deep neural network model.
6. The method according to claim 1 , wherein the deep neural network model comprises one or more of disease classification, disease area detection, or disease area segmentation.
7. An apparatus for analyzing a medical image, the apparatus comprising:
at least one processor; and
a memory storing instructions, the instructions when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising:
acquiring medical image data;
generating multi-scale decision sample data based on the medical image data; and
inputting the multi-scale decision sample data into a deep neural network model to obtain an auxiliary diagnosis data of the medical image, wherein the deep neural network model is trained using multi-scale training sample data according to a consistency principle between the multi-scale training sample data and an output result, wherein the deep neural network model is trained by:
acquiring full-field pathological image data;
generating the multi-scale training sample data based on the full-field pathological image data; and
training the deep neural network model according to a consistency principle between the multi-scale training sample data and the output result of the deep neural network model using the multi-scale training sample data, and
wherein the generating the multi-scale training sample data based on the full-field pathological image data comprises:
removing a non-effective pathological tissue area in the full-field pathological image data to obtain an effective pathological tissue area; and
generating the multi-scale training sample data based on the effective pathological tissue area.
8. The apparatus according to claim 7 , wherein the generating the multi-scale training sample data based on the full-field pathological image data comprises:
determining a positional correspondence relationship between the multi-scale training sample data using a multi-scale alignment algorithm.
9. The apparatus according to claim 7 , wherein the training the deep neural network model using the multi-scale training sample data comprises:
training the deep neural network model respectively using training sample data of different scales in the multi-scale training sample data; or
training the deep neural network model jointly using the training sample data of different scales in the multi-scale training sample data.
10. The apparatus according to claim 7 , wherein the training the deep neural network model using the multi-scale training sample data further comprises:
establishing a connection between training sample data of different scales in the multi-scale training sample data using topology or dynamic coding; and
training the deep neural network model based on the connection between the training sample data of different scales.
11. The apparatus according to claim 7 , wherein the training the deep neural network model using the multi-scale training sample data comprises at least one of:
selecting sample data with a score lower than a predetermined threshold from the multi-scale training sample data for back propagation, to update parameters of the deep neural network model; or
selecting sample data expired in at least one of training or detection from the multi-scale training sample data for hard example mining, to update parameters of the deep neural network model.
12. The apparatus according to claim 7 , wherein the deep neural network model comprises one or more of disease classification, disease area detection, or disease area segmentation.
13. A non-transitory computer readable storage medium, storing a computer program thereon, wherein the computer program, when executed by a processor, causes the processor to perform operations, the operations comprising:
acquiring medical image data;
generating multi-scale decision sample data based on the medical image data; and
inputting the multi-scale decision sample data into a deep neural network model to obtain auxiliary diagnosis data of the medical image, wherein the deep neural network model is trained using multi-scale training sample data according to a consistency principle between the multi-scale training sample data and an output result, wherein the deep neural network model is trained by:
acquiring full-field pathological image data;
generating the multi-scale training sample data based on the full-field pathological image data; and
training the deep neural network model according to a consistency principle between the multi-scale training sample data and an output result of the deep neural network model using the multi-scale training sample data, and
wherein the generating the multi-scale training sample data based on the full-field pathological image data comprises:
removing a non-effective pathological tissue area in the full-field pathological image data to obtain an effective pathological tissue area; and
generating the multi-scale training sample data based on the effective pathological tissue area.