Method for predicting epidermal growth factor receptor mutation in brain metastases from non-small cell lung cancer

Abstract

This invention discloses a method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer. The method includes an image preprocessing step: fusing a magnetic resonance imaging (MRI) image and a computed tomography (CT) image to obtain a tumor volume; a 3D model image processing step: taking a first distance outward from an edge of the tumor volume and cropping a foreground bounding box, adjusting the foreground bounding box to a first matrix size; a 2D model image processing step: similar to the 3D model image processing step, adjusting the foreground bounding box to a second matrix size, cropping it into multiple slice images to generate multiple 2D image sets; and a model inference step: inferring the predicted value of each 2D image set, then combining each predicted value according to a predetermined weight set and summing them to obtain a final predicted value of the tumor volume.

Claims

1. A method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer, comprising: Image preprocessing step: Rigid registration is performed on a magnetic resonance imaging (MRI) image and a computed tomography (CT) image to fuse a tumor volume. The image intensity of the tumor volume is standardized to ensure that the intensity value of the magnetic resonance signal within a 5 mm range on the surface of the tumor volume conforms to a standard intensity range of 0 to 1024. 3D model image processing step: Based on an edge of the tumor volume, a first distance is taken outward and a foreground frame is cropped. Then, the spatial size of the foreground frame is adjusted to a first matrix size that is a cube. A 2D model image processing step: Based on the edge of the tumor volume, take the first distance outward and cut out the foreground box. Then, adjust the spatial size of the foreground box to a second matrix size of a cuboid. Cut into multiple slice images according to a predetermined direction. Then, merge every three slice images to form a three-channel image to generate multiple 2D image sets. A model inference step: Infer the prediction values ​​of each of the multiple 2D image sets respectively. Then, according to a predetermined weight set, combine the prediction values ​​of each of the multiple 2D image sets and add them together to obtain a final prediction value for the tumor volume. The edge of the tumor volume is created by isotropically expanding 3 mm and isotropically contracting 2 mm around the surface of the tumor volume to highlight the radiographic features of the surrounding area of ​​the tumor.

2. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, The image preprocessing step also includes an N4 bias correction algorithm prior to standardization to eliminate low-frequency intensity variations in the tumor volume.

3. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 2, wherein, This image preprocessing step, following the standardization, also includes resampling the tumor volume 1,000 times.

4. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, The first distance is 4mm.

5. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, The size of the first matrix is ​​64 x 64 x 64.

6. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, The size of the second matrix is ​​64 x 64 x 33.

7. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, The 3D model image processing step also includes performing an image enhancement, which at least includes using a random affine transformation to perform rotation and scaling with a specific range of rotation and scaling.

8. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 7, wherein, The 3D model image processing steps also include Z-score normalization after image enhancement to ensure that the intensity distribution of the processed tumor volume is consistent.

9. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, Each of the three-channel images is an R, G, B channel image.

10. The method for predicting epidermal growth factor receptor mutations in brain metastases of non-small cell lung cancer as described in claim 1, wherein, In the 2D model image processing step, the image is cropped into 33 slices according to the predetermined direction of z-axis, and then every 3 slices are merged to form a 2D image set, resulting in a total of 11 2D image sets. In the model inference step, the predicted values ​​of the 11 2D image sets are inferred respectively. Then, the predicted values ​​of the 11 2D image sets are added together according to the predetermined weight set, which includes 11 weights of 0, 0, 0, 1, 2, 5, 2, 1, 0, 0, 0, to obtain the final predicted value for the tumor volume.

Images