Research on Acetabular Labrum Injury Based on MR: Multi-angle Deep Learning Model

Overview

The purpose of this study is to study the MRI images of acetabular labrum injury by deep learning method, and try to establish a combination model of axial and coronal serial images, and compare with the diagnostic accuracy of radiologists, to establish a deep learning method for accurate identification and classification of acetabular labrum injury.

1. Detection of acetabular labrum images based on CNN: axial and coronal T2-fs images were used, and all images were corrected and standardized. CNN is applied to recognize and learn the images with acetabular labrum to select the images with acetabular labrum structure from the complete sequence and delete the images without acetabular labrum structure. All the data are divided into a training set (70%, 30% in training set as verification set), and the remaining 30% as a test set to evaluate the accuracy of model recognition. Enter the obtained results into the next step. 2. Recognition and segmentation of acetabular labrum based on Densenet: using Densenet to recognize and learn the acetabular labrum from the selected images. LabelImg was used to locate the acetabular labrum coordinates manually and then input them into Python for recognition learning. All the data were divided into a training set (70%, and then 30% in the training set was selected as the verification set), and the remaining 30% was used as the test set to evaluate the accuracy of model recognition. After identifying the labrum structure, the labrum structure is locally cut and enlarged to remove the redundant information. Finally, input the result to the next step. 3. Identification and grading of acetabular labrum injury based on 3D-CNN: the input data were identified and graded by the 3D-CNN model. 3D-CNN is divided into eight layers: input layer, hard wire layer H1, convolution layer C2, downsampling layer S3, convolution layer C4, downsampling layer S5, convolution layer C6 and output layer. 3D-CNN constructs a cube by stacking multiple consecutive frames and then uses a 3D convolution kernel in the cube. Through this structure, the feature images in the convolution layer will be connected with multiple adjacent frames in the previous layer to realize the information acquisition of continuous images. Similarly, the data were divided into a training set (70%, and then 30% of the training set was selected as the verification set), and the remaining 30% was used as the test set to evaluate the classification accuracy to identify the injury of the labrum and classify the cases with injury. 4. Combination model: according to the above process (1-3), after the models are established for the axial and coronal images respectively, according to the output characteristics of the CNN classification model, the probabilities of different grades are predicted before the output results, and the output results are based on these probabilities to select the expression form of the maximum possible probability. Our combination model averages the probabilities of these different classifications, calculates the final prediction probability, and then obtains the final model. The test set of the third step (including the mixed data of axial and coronal images) was used to verify the model. 5. Comparison of radiologists and deep learning: List the test set cases in the above step 3 and ask two MSK professional radiologists to classify whether there is damage and the degree of damage, and compared with the results with artificial diagnosis (both doctors read the images independently and obtained the diagnosis results by simulating the normal state of writing the report without any prompt). Finally, the accuracy of artificial diagnosis was compared with that of the combination model obtained in the fourth step.

Conditions

Interventions

Eligibility

Locations (1)

Outcomes