Multi-task learning for chest X-ray abnormality classification

1. A system for multi-abnormality classification based on chest X-ray images, the system comprising:

an imaging database configured to store a chest X-ray image;

a normalization module configured to process the chest X-ray image to remove image characteristic variability due to acquisition factors;

a multi-task network configured to receive the chest X-ray image and output abnormality classification scores for a plurality of abnormalities, wherein the multi-task network comprises an encoder network and a decoder network, wherein the encoder network comprises a plurality of layers of densely connected blocks followed by a global average pooling layer to predict the abnormality classification scores for the chest X-ray image;

a confidence module configured to map each of the abnormality classification scores to a discrete classification as a function of a learned score threshold and a discrete confidence category; and

an interface configured to output the discrete classifications.

2. The system of claim 1 , further comprising:

an imaging system configured to acquire the chest X-ray image.

3. The system of claim 1 , wherein the normalization module is configured to adjust brightness and contrast via a linear transformation of image intensities of the chest X-ray image.

4. The system of claim 1 , wherein the multi-task network outputs abnormality classification scores of abnormalities including at least two of granuloma, infiltrate, nodule, scaring, effusion, atelectasis, bone or soft tissue lesion, fibrosis, cardiac abnormality, mass, pneumothorax, COPD, consolidation, pleural thickening, cardiomegaly, emphysema, edema, pneumonia, hilar abnormality, or hernias.

5. The system of claim 1 , wherein the decoder network comprises up-sampling, convolutional, and nonlinear layers and is configured to generate predicted segmented masks of anatomical structures for the chest X-ray image.

6. The system of claim 1 , wherein a global loss value used for optimizing internal parameters of the multi-task network is calculated as a combination of an abnormality classification loss value, a segmentation loss value, and a spatial location loss value.

7. The system of claim 1 , wherein the learned score threshold is calculated as a function of a multi-user observer study to yield a 1:1 ratio of false positives and false negatives.

8. The system of claim 1 , wherein the confidence module is configured to map the abnormality classification scores as a function of a multi-user observer study to span a range such that a ratio of correct to incorrect classifications for each abnormality reaches a prespecified value.

9. The system of claim 1 , further comprising:

an analysis module configured to recommend a procedure as a function of the discrete classification of abnormalities.

10. A method for training a multi-task network for classification of different abnormalities, classification of locations of the different abnormalities, and segmentation of lung lobes and heart, the method comprising:

acquiring training data comprising a plurality of chest X-ray images and annotations;

normalizing the chest X-ray images to remove image characteristic variability due to acquisition factors;

inputting the normalized chest X-ray images into the multi-task network;

outputting, by the multi-task network, abnormality classification scores and a segmented mask;

comparing the abnormality classification scores and the segmented mask against the annotations from the training data;

adjusting weights in the multi-task network as a function of the comparison;

repeating inputting, outputting, comparing, and adjusting for a predetermined number of iterations; and

outputting a trained multi-task network.

11. The method of claim 10 , wherein normalizing comprises adjusting a brightness and contrast of the chest X-ray images via a linear transformation of image intensities of the chest X-ray images.

12. The method of claim 10 , wherein the multi-task network comprises an encoder decoder network.

13. The method of claim 12 , wherein the encoder network comprises a plurality of layers of densely connected blocks followed by a global average pooling layer to predict the abnormality classification scores for the chest X-ray image.

14. The method of claim 12 , wherein the decoder network comprises up-sampling, convolutional, and nonlinear layers and is configured to generate predicted segmented masks of anatomical structures for the chest X-ray image.

15. The method of claim 10 , wherein comparing comprises calculating a global loss value as a combination of an abnormality classification loss value and a segmentation loss value.

16. A non-transitory computer implemented storage medium, including machine-readable instructions stored therein, that when executed by at least one processor, cause the processor to:

acquire a chest X-ray image of a patient;

normalize the chest X-ray image;

input the normalized chest X-ray image into a multi-task machine learnt network optimized to identify abnormalities in chest X-ray images;

receive, from the multi-task machine learnt network, abnormality classification scores for the chest X-ray image;

map each of the abnormality classification scores to a discrete classification of abnormalities as a function of a multi-user observer study so that the discrete classification of abnormalities span a range such that a ratio of correct to incorrect classifications for each abnormality reaches a prespecified value; and

output the discrete classifications of abnormalities.

17. The non-transitory computer implemented storage medium of claim 16 , further comprising machine-readable instructions that when executed by at least one processor, cause the processor to:

recommend a medical procedure as a function of the discrete classification of abnormalities.

18. The non-transitory computer implemented storage medium of claim 16 , wherein the machine-readable instructions to normalize the chest X-ray image comprise instructions to adjust brightness and contrast via a linear transformation of image intensities of the chest X-ray image.