Automatic detection of lesions in medical images using 2D and 3D deep learning networks
Systems and methods for automatic segmentation of lesions from a 3D input medical image are provided. A 3D input medical image depicting one or more lesions is received. The one or more lesions are segmented from one or more 2D slices extracted from the 3D input medical image using a trained 2D segmentation network. 2D features are extracted from results of the segmentation of the one or more lesions from the one or more 2D slices. The one or more lesions are segmented from a 3D patch extracted from the 3D input medical image using a trained 3D segmentation network. 3D features are extracted from results of the segmentation of the one or more lesions from the 3D patch. The extracted 2D features and the extracted 3D features are fused to generate final segmentation results. The final segmentation results are output.
1. A method comprising:
receiving a 3D input medical image depicting one or more lesions;
segmenting the one or more lesions from one or more 2D slices extracted from the 3D input medical image using a trained 2D segmentation network;
extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices;
segmenting the one or more lesions from a 3D patch extracted from the 3D input medical image using a trained 3D segmentation network;
extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch;
fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results, wherein the trained machine learning based fusion model is trained based on 1) probability values extracted from results of a segmentation of 2D training slices, 2) probability values extracted from results of a segmentation of 3D training images, and 3) ground truth annotations of the probability values extracted from the results of the segmentation of 2D training slices and the 3D training images; and
outputting the final segmentation results.
2. The method of claim 1 , wherein:
extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices comprises extracting the first probability values from a prediction map output from the trained 2D segmentation network, and
extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch comprises extracting the second probability values from a prediction map output from the trained 3D segmentation network.
3. The method of claim 1 , further comprising:
training the trained machine learning based fusion model by ensemble learning.
4. The method of claim 1 , wherein fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
concatenating the first probability values and the second probability values.
5. The method of claim 1 , wherein fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
combining the first probability values and the second probability values based on a weighted vote.
6. The method of claim 1 , wherein fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
fusing the first probability values and the second probability values to generate a final prediction map representing a voxelwise score that a voxel depicts a lesion.
7. The method of claim 1 , wherein the one or more 2D slices comprises a plurality of orientation images.
8. The method of claim 1 , wherein the 3D input medical image comprises an MM (magnetic resonance imaging) image of a brain of a patient and the one or more lesions comprises one or more brain metastases.
9. An apparatus comprising:
means for receiving a 3D input medical image depicting one or more lesions;
means for segmenting the one or more lesions from one or more 2D slices extracted from the 3D input medical image using a trained 2D segmentation network;
means for extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices;
means for segmenting the one or more lesions from a 3D patch extracted from the 3D input medical image using a trained 3D segmentation network;
means for extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch;
means for fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results, wherein the trained machine learning based fusion model is trained based on 1) probability values extracted from results of a segmentation of 2D training slices, 2) probability values extracted from results of a segmentation of 3D training images, and 3) ground truth annotations of the probability values extracted from the results of the segmentation of 2D training slices and the 3D training images; and
means for outputting the final segmentation results.
10. The apparatus of claim 9 , wherein:
the means for extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices comprises means for extracting the first probability values from a prediction map output from the trained 2D segmentation network, and
the means for extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch comprises means for extracting the second probability values from a prediction map output from the trained 3D segmentation network.
11. The apparatus of claim 9 , further comprising:
means for training the trained machine learning based fusion model by ensemble learning.
12. The apparatus of claim 9 , wherein the means for fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
means for concatenating the first probability values and the second probability values.
13. A non-transitory computer readable medium storing computer program instructions, the computer program instructions when executed by a processor cause the processor to perform operations comprising:
receiving a 3D input medical image depicting one or more lesions;
segmenting the one or more lesions from one or more 2D slices extracted from the 3D input medical image using a trained 2D segmentation network;
extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices;
segmenting the one or more lesions from a 3D patch extracted from the 3D input medical image using a trained 3D segmentation network;
extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch;
fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results, wherein the trained machine learning based fusion model is trained based on 1) probability values extracted from results of a segmentation of 2D training slices, 2) probability values extracted from results of a segmentation of 3D training images, and 3) ground truth annotations of the probability values extracted from the results of the segmentation of 2D training slices and the 3D training images; and
outputting the final segmentation results.
14. The non-transitory computer readable medium of claim 13 , wherein:
extracting first probability values from results of the segmentation of the one or more lesions from the one or more 2D slices comprises extracting the first probability values from a prediction map output from the trained 2D segmentation network, and
extracting second probability values from results of the segmentation of the one or more lesions from the 3D patch comprises extracting the second probability values from a prediction map output from the trained 3D segmentation network.
15. The non-transitory computer readable medium of claim 13 , wherein fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
combining the first probability values and the second probability values based on a weighted vote.
16. The non-transitory computer readable medium of claim 13 , wherein fusing the first probability values and the second probability values using a trained machine learning based fusion model to generate final segmentation results comprises:
fusing the first probability values and the second probability values to generate a final prediction map representing a voxelwise score that a voxel depicts a lesion.
17. The non-transitory computer readable medium of claim 13 , wherein the one or more 2D slices comprises a plurality of orientation images.
18. The non-transitory computer readable medium of claim 13 , wherein the 3D input medical image comprises an MM (magnetic resonance imaging) image of a brain of a patient and the one or more lesions comprises one or more brain metastases.