System and method for breast cancer detection in X-ray images
For breast cancer detection with an x-ray scanner, a cascade of multiple classifiers is trained or used. One or more of the classifiers uses a deep-learnt network trained on non-x-ray data, at least initially, to extract features. Alternatively or additionally, one or more of the classifiers is trained using classification of patches rather than pixels and/or classification with regression to create additional cancer-positive partial samples.
1. A method for breast cancer detection with an x-ray scanner, scanner, the method comprising:
scanning a patient with the x-ray scanner, the scanning providing a frame of data representing breast tissue in the patient;
classifying patches of the frame of data as of interest or not based on intensity;
determining a probability of breast cancer for the patches that are of interest using a deep learnt regression at least partially trained on non-x-ray data, the probability not determined for the patches not of interest, the determining comprising inputting the patches that are of interest into the deep learnt regression, the deep regression outputting feature values;
classifying the patches with the probability over a threshold using a machine-learnt classifier and not classifying the patches with the probability below the threshold; and
displaying an indication of breast cancer for the patient based on an output of the classifying using the machine-learnt classifier.
2. The method of claim 1 wherein scanning comprises acquiring a mammogram.
3. The method of claim 1 wherein classifying the patches based on intensity comprises dividing the frame of data into the patches, calculating mean intensity for each of the patches, and classifying based on the mean intensity.
4. The method of claim 1 wherein determining the probability comprises using the regression model based on intermediate features assigned to first locations adjacent to second locations with higher probabilities and to third locations with lower probabilities.
5. The method of claim 1 wherein classifying using the machine-learnt classifier comprises classifying with the machine-learnt classifier learned with regression.
6. The method of claim 1 wherein classifying using the machine-learnt classifier comprises sampling locations in each patch, classifying with the machine-learnt classifier for kernels centered at the sampling locations, combining results of the classifying for the kernels, and providing the output as a function of the combined results.
7. The method of claim 1 wherein displaying comprises displaying an image from the frame of data with highlighting of locations for the patches classified using the machine-learnt classifier as having breast cancer.
8. The method of claim 1 wherein displaying comprises displaying an x-ray image from the frame of data when any of the patches classified using the machine-learnt classifier is classified as having breast cancer.
9. A method for machine training a classifier for breast cancer detection, the method comprising:
acquiring a set of x-ray images with ground truth labels for first locations of breast cancer;
assigning adjacent locations to the first locations with regressed labels of breast cancer and second locations spaced from the first locations as ground truth labels of no breast cancer, the regressed labels being based on a distance transform and reduction from the ground truth labels; and
machine training a cascade of classifiers, at least one of the cascade of classifiers trained using the ground truth labels of breast cancer for the first locations, the regressed labels for the adjacent locations, and the ground truth labels of no breast cancer for the second locations.
10. The method of claim 9 wherein assigning comprises assigning the regressed labels as a function of distance from the ground truth labels of breast cancer for the first locations, greater distance having greater regression.
11. The method of claim 9 wherein the machine training comprises training the cascade of the classifiers as first and second classifiers, the first classifier trained using the ground truth labels of breast cancer and no breast cancer, the second classifier trained using the regressed labels.
12. The method of claim 9 wherein machine training comprises machine training with feature values output by a deep learnt network trained on image data other than medical imaging data.
13. The method of claim 12 wherein machine training with the deep learnt network comprises machine training with the deep learnt network trained with the x-ray images.
14. The method of claim 12 wherein machine training comprises machine training with a regression analysis.
15. The method of claim 12 wherein machine training comprises machine training with a support vector machine operable with the regressed labels.
16. A method for breast cancer detection with an x-ray scanner, the method comprising:
scanning a patient with the x-ray scanner, the scanning providing a frame of data representing breast tissue in the patient;
determining a probability of breast cancer for the patient with a deep learnt network initially trained on image data for objects other than breast tissue, the probability determined with feature values output by the deep learnt network and with a learnt regression function;
displaying an indication of breast cancer for the patient based on the probability.
17. The method of claim 16 wherein determining the probability is a first classifier phase of a cascade of classifier phases.
18. The method of claim 16 wherein determining the probability comprises determining with the deep learnt network trained on image data from optical photographs of the objects in a non-medical environment, and with the deep learnt network incrementally trained with x-ray images.