Method and device for detecting and locating lesion in medical image, equipment and storage medium

A computer equipment, comprising:

• a memory, a processor, and a computer readable instruction stored in the memory and capable of running on the processor, wherein the processor, when executing the computer readable instruction, implements:
• obtaining a target medical image of a lesion to be detected;
• inputting the target medical image into a pre-trained deep learning model to obtain a target sequence output from the deep learning model, wherein each element in the target sequence is a first confidence corresponding to each preset lesion type, wherein the first confidence represents a probability that the target medical image belongs to a corresponding preset lesion type, wherein the deep learning model is obtained by pre-training a medical image sample corresponding to each preset lesion type, and wherein each medical image sample is marked with a lesion type comprised in the target medical image;
• after inputting the target medical image into the deep learning model, extracting a first feature map output from a last convolution layer in the deep learning model;
• extracting a weight value of each network unit corresponding to each preset lesion type in a fully connected layer of the deep learning model;
• for each preset lesion type, calculating a fusion feature map corresponding to each preset lesion type according to the first feature map and the weight value corresponding to each preset lesion type;
• resampling the fusion feature map corresponding to each preset lesion type to a size of the target medical image to generate a generic activation map corresponding to each preset lesion type;
• determining a maximum connected area in the generic activation map corresponding to each preset lesion type, and creating a mark border surrounding the maximum connected area, wherein the maximum connected area refers to a connected area into which a point whose pixel value exceeds a specified pixel threshold falls in the generic activation map;
• and
• adding a mark border corresponding to each preset lesion type to the target medical image as a location result of the lesion to be detected, wherein
• the processor, when executing the computer readable instruction, further implements after inputting the target medical image into the pre-trained deep learning model to obtain the target sequence output from the deep learning model:
• obtaining a preset confidence threshold corresponding to each preset lesion type;
• comparing each element in the target sequence with the corresponding preset confidence threshold to determine an element greater than or equal to the preset confidence threshold in the target sequence, denoting it as a target element;
• and
• determining a prediction result of the preset lesion type corresponding to each target element on the target medical image to be positive, and determining the prediction result of the preset lesion types corresponding other elements on the target medical image to be negative, wherein the other elements refer to the elements in the target sequence other than the target element;
• and
• the preset confidence threshold corresponding to each preset lesion type is predetermined as follows:
• obtaining each medical image that is collected in advance for verification and denoting it as each verification sample;
• for each verification sample, inputting each verification sample into the pre-trained deep learning model, and obtaining a verification sequence corresponding to each verification sample output from the deep learning model, wherein each element in the verification sequence is a third confidence corresponding to each preset lesion type, and wherein the third confidence represents a probability that each verification sample belongs to a corresponding preset lesion type;
• for each preset lesion type, extracting the third confidence corresponding to each preset lesion type in the verification sequence corresponding to each verification sample to obtain each lesion confidence;
• for each preset lesion type, according to each lesion confidence, calculating a Yoden index with each lesion confidence as a threshold to obtain an index result corresponding to each lesion confidence;
• and
• determining the lesion confidence corresponding to the maximum index result as the preset confidence threshold corresponding to each preset lesion type, wherein the maximum index result refers to the maximum of the index results corresponding to the lesion confidences.

A non-transitory computer readable storage medium that stores a computer readable instruction, wherein the computer readable instruction, when executed by one or more processors, enables the one or more processors to perform:

• obtaining a target medical image of a lesion to be detected;
• inputting the target medical image into a pre-trained deep learning model to obtain a target sequence output from the deep learning model, wherein each element in the target sequence is a first confidence corresponding to each preset lesion type, wherein the first confidence represents a probability that the target medical image belongs to a corresponding preset lesion type, wherein the deep learning model is obtained by pre-training a medical image sample corresponding to each preset lesion type, and wherein each medical image sample is marked with a lesion type comprised in the target medical image;
• after inputting the target medical image into the deep learning model, extracting a first feature map output from a last convolution layer in the deep learning model;
• extracting a weight value of each network unit corresponding to each preset lesion type in a fully connected layer of the deep learning model;
• for each preset lesion type, calculating a fusion feature map corresponding to each preset lesion type according to the first feature map and the weight value corresponding to each preset lesion type;
• resampling the fusion feature map corresponding to each preset lesion type to a size of the target medical image to generate a generic activation map corresponding to each preset lesion type;
• determining a maximum connected area in the generic activation map corresponding to each preset lesion type, and creating a mark border surrounding the maximum connected area, wherein the maximum connected area refers to a connected area into which a point whose pixel value exceeds a specified pixel threshold falls in the generic activation map;
• and
• adding a mark border corresponding to each preset lesion type to the target medical image as a location result of the lesion to be detected, wherein
• the computer readable instruction, when executed by the one or more processors, enables the one or more processors to further perform:
• after inputting the target medical image into the pre-trained deep learning model to obtain the target sequence output from the deep learning model:
• obtaining a preset confidence threshold corresponding to each preset lesion type;
• comparing each element in the target sequence with the corresponding preset confidence threshold to determine an element greater than or equal to the preset confidence threshold in the target sequence, and denoting it as a target element;
• and
• determining a prediction result of the preset lesion type corresponding to each target element on the target medical image to be positive, and determining the prediction result of the preset lesion types corresponding other elements on the target medical image to be negative, wherein the other elements refer to the elements in the target sequence other than the target element;
• and
• the preset confidence threshold corresponding to each preset lesion type is predetermined as follows:
• obtaining each medical image that is collected in advance for verification and denoting it as each verification sample;
• for each verification sample, inputting each verification sample into the pre-trained deep learning model, and obtaining a verification sequence corresponding to each verification sample output from the deep learning model, wherein each element in the verification sequence is a third confidence corresponding to each preset lesion type, and wherein the third confidence represents a probability that each verification sample belongs to a corresponding preset lesion type;
• for each preset lesion type, extracting the third confidence corresponding to each preset lesion type in the verification sequence corresponding to each verification sample to obtain each lesion confidence;
• for each preset lesion type, according to each lesion confidence, calculating a Yoden index with each lesion confidence as a threshold to obtain an index result corresponding to each lesion confidence;
• and
• determining the lesion confidence corresponding to the maximum index result as the preset confidence threshold corresponding to each preset lesion type, wherein the maximum index result refers to the maximum of the index results corresponding to the lesion confidences.