Evaluating and executing method for model and electronic device

Abstract

This invention provides a method and electronic device for evaluating and executing a model, applicable to medical settings, and capable of improving the accuracy and stability of the selected model. The method includes the following steps: Calculating a weighted sum of multiple indicator values ​​for each of multiple candidate models based on multiple weight values; Selecting one candidate model as the output model based on the multiple weighted sums; Executing the output model to infer meaning from medical images to produce inference results, providing symptom structures interpreted based on the inference results.

Claims

1. A method for evaluating and implementing a model, comprising: Through a processor, a weighted sum of values ​​for each of the multiple candidate models is calculated based on multiple weight values. The processor selects one of the candidate models as an output model based on the weighted sum values; and the processor executes the output model to infer a medical image to produce an inference result, providing a symptom structure interpreted based on the inference result, including: the processor performs a standardization process on the medical image to produce a standardized image. And through the processor, a dual inference is performed on the medical image and the standardized image based on multiple inference weight values ​​associated with the symptom structure to produce the inference result, wherein the inference weight values ​​are associated with the target disease information corresponding to the symptom structure, the symptom structure including a retinal nerve fiber layer (RNFL), and the target disease information is the clinical usage preference of the symptom structure.

2. The evaluation and execution method as described in claim 1 further includes: The processor calculates the following metrics for each candidate model based on a validation set: Area Under Curve (AUC), Sensitivity, Specificity, F1 Score, and Normalized Entropy.

3. The evaluation and execution method as described in claim 2, wherein the step of calculating the weighted sum of the candidate models for the index values ​​based on the weight values ​​by the processor includes: The processor calculates a first product of the AUC value and a first weight value. The processor calculates an average of the sensitivity value and the specificity value, and calculates a second product of the average value and a second weight value; the processor calculates a third product of the F1 score and a third weight value; the processor calculates a fourth product of the normalized entropy and a fourth weight value. And through the processor, the first product value, the second product value, the third product value, and the fourth product value are summed to generate the weighted total value.

4. The evaluation and execution method as described in claim 2, wherein the validation set includes multiple medical images belonging to different pathological categories.

5. The evaluation and execution method as described in claim 1 further includes: The processor sets weight values ​​based on the target disease information corresponding to the symptom structure.

6. The evaluation and execution method as described in claim 1 further includes: The processor sorts the candidate models according to the weighted sum values. The processor then selects the candidate model with the largest weighted sum value as the model to be tested.

7. The evaluation and execution method as described in claim 6 further includes: The processor executes the model under test to infer a test set to generate a test inference result, wherein the test set includes multiple test medical images from different models of equipment; and the processor determines whether to use the model under test as the output model based on the test inference result and preset performance data.

8. The evaluation and execution method as described in claim 1 further includes: Through the processor, based on the target disease information corresponding to the symptom structure, a first inference weight value and a second inference weight value are set among these inference weight values, wherein the sum of the first inference weight value and the second inference weight value is equal to 1, and the first inference weight value is greater than the second inference weight value.

9. The evaluation and execution method as described in claim 8, wherein the inference result includes a first inference result corresponding to the standardized image based on the first inference weight value, and a weighted summation result of a second inference result corresponding to the medical image based on the second inference weight value.

10. An electronic device comprising: One memory is used to store multiple candidate models and multiple weight values; And a processor coupled to the memory for performing the evaluation and execution methods as described in claim 1.

Images