Lesion detection and segmentation
Mechanisms are provided for detecting lesions in diffusion weighted imaging (DWI) images. The mechanisms receive a first set of DWI images corresponding to a anatomical structure, from medical imaging computer system(s). The first set of DWI images comprises a plurality of DWI images having at least two different b-values. The mechanisms generate a second set of DWI images from the first set of DWI images based on at least one predetermined criterion. The second set of DWI images comprises different DWI images having different b-values. The mechanisms extract feature data from the second set of DWI images, input the feature data into at least one computer neural network, and generate an output from the neural network(s) comprising at least one of a lesion classification or a lesion mask based on results of processing, by the neural network(s), of the feature data extracted from the second set of DWI images.
1. A method for detecting lesions in diffusion weighted imaging (DWI) images, the method comprising:
receiving a first set of DWI images corresponding to a given anatomical structure, from at least one medical imaging computer system, wherein the first set of DWI images comprises a plurality of DWI images having at least two different b-values;
generating a second set of DWI images from the first set of DWI images based on at least one predetermined criterion, wherein the second set of DWI images comprises different DWI images having different b-values;
extracting feature data from the second set of DWI images;
inputting the feature data into at least one computer neural network; and
generating an output from the computer neural network comprising at least one of a lesion classification or a lesion mask based on results of processing, by the at least one computer neural network, of the feature data extracted from the second set of DWI images.
2. The method of claim 1 , wherein generating the second set of DWI images from the first set of DWI images based on the at least one predetermined criterion comprises:
selecting a first subset of DWI images, from the first set of DWI images, having a lowest b-value relative to other b-values of other DWI images in the first set of DWI images;
selecting a second subset of DWI images, from the first set of DWI images, having a highest b-value relative to other b-values of other DWI images in the first set of DWI images; and
combining the first subset and second subset to form the second set of DWI images.
3. The method of claim 2 , wherein generating the second set of DWI images further comprises generating an apparent diffusion coefficient (ADC) map from the first set of DWI images and combining the ADC map with the first subset and second subset to form the second set of DWI images.
4. The method of claim 1 , wherein generating the second set of DWI images comprises generating a third set of DWI images from the first set of DWI images having a corresponding b-value that is not represented in the b-values associated with DWI images in the first set of DWI images.
5. The method of claim 4 , wherein generating the third set of DWI images from the first set of DWI images comprises at least one of interpolating or extrapolating a number of DWI images having different b-value from the b-values associated with DWI images of the first set of DWI images.
6. The method of claim 4 , wherein generating the third set of DWI images from the first set of DWI images comprises generating a number of DWI images having a higher b-value than b-values associated with DWI images in the first set of DWI images.
7. The method of claim 1 , wherein extracting feature data from the second set of DWI images comprises processing the second set of DWI images as input to at least one trained neural network that flattens data corresponding to the set of DWI images and applies positional encoding to the flattened data to generate the feature data.
8. The method of claim 1 , wherein the at least one computer neural network comprises a multi-head transformer that comprises an internal transformer, having an encoder and a decoder, and a plurality of heads coupled to the internal transformer that receive outputs from the decoder.
9. The method of claim 8 , wherein the plurality of heads comprise a classification head that processes a first output from the decoder to generate a classification of whether the feature data input to the internal transformer indicates the presence of a lesion, and a segmentation head that processes a second output from the decoder to generate a lesion mask based on the feature data input to the internal transformer.
10. The method of claim 8 , wherein the at least one computer neural network is trained through a machine learning training operation based on a combined cross-head loss function that combines a weighted classification loss associated with the classification head and a weighted segmentation loss associated with the segmentation head.
11. A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to be configured to perform lesion detection in diffusion weighted imaging (DWI) images at least by:
receiving a first set of DWI images corresponding to a given anatomical structure, from at least one medical imaging computer system, wherein the first set of DWI images comprises a plurality of DWI images having at least two different b-values;
generating a second set of DWI images from the first set of DWI images based on at least one predetermined criterion, wherein the second set of DWI images comprises different DWI images having different b-values;
extracting feature data from the second set of DWI images;
inputting the feature data into at least one computer neural network; and
generating an output from the at least one computer neural network comprising at least one of a lesion classification or a lesion mask based on results of processing, by the at least one computer neural network, of the feature data extracted from the second set of DWI images.
12. The computer program product of claim 11 , wherein generating the second set of DWI images from the first set of DWI images based on the at least one predetermined criterion comprises:
selecting a first subset of DWI images, from the first set of DWI images, having a lowest b-value relative to other b-values of other DWI images in the first set of DWI images;
selecting a second subset of DWI images, from the first set of DWI images, having a highest b-value relative to other b-values of other DWI images in the first set of DWI images; and
combining the first subset and second subset to form the second set of DWI images.
13. The computer program product of claim 12 , wherein generating the second set of DWI images further comprises generating an apparent diffusion coefficient (ADC) map from the first set of DWI images and combining the ADC map with the first subset and second subset to form the second set of DWI images.
14. The computer program product of claim 11 , wherein generating the second set of DWI images comprises generating a third set of DWI images from the first set of DWI images having a corresponding b-value that is not represented in the b-values associated with DWI images in the first set of DWI images.
15. The computer program product of claim 14 , wherein generating the third set of DWI images from the first set of DWI images comprises at least one of interpolating or extrapolating a number of DWI images having different b-value from the b-values associated with DWI images of the first set of DWI images.
16. The computer program product of claim 14 , wherein generating the third set of DWI images from the first set of DWI images comprises generating a number of DWI images having a higher b-value than b-values associated with DWI images in the first set of DWI images.
17. The computer program product of claim 11 , wherein extracting feature data from the second set of DWI images comprises processing the second set of DWI images as input to at least one trained neural network that flattens data corresponding to the set of DWI images and applies positional encoding to the flattened data to generate the feature data.
18. The computer program product of claim 11 , wherein the at least one computer neural network comprises a multi-head transformer that comprises an internal transformer, having an encoder and a decoder, and a plurality of heads coupled to the internal transformer that receive outputs from the decoder.
19. The computer program product of claim 18 , wherein the plurality of heads comprise a classification head that processes a first output from the decoder to generate a classification of whether the feature data input to the internal transformer indicates the presence of a lesion, and a segmentation head that processes a second output from the decoder to generate a lesion mask based on the feature data input to the internal transformer.
20. An apparatus comprising:
at least one processor; and
at least one memory coupled to the at least one processor, wherein the at least one memory comprises instructions which, when executed by the at least one processor, causes the at least one processor to be configured to perform lesion detection in diffusion weighted imaging (DWI) images at least by:
receiving a first set of DWI images corresponding to a given anatomical structure, from at least one medical imaging computer system, wherein the first set of DWI images comprises a plurality of DWI images having at least two different b-values;
generating a second set of DWI images from the first set of DWI images based on at least one predetermined criterion, wherein the second set of DWI images comprises different DWI images having different b-values;
extracting feature data from the second set of DWI images;
inputting the feature data into at least one computer neural network; and
generating an output from the at least one computer neural network comprising at least one of a lesion classification or a lesion mask based on results of processing, by the at least one computer neural network, of the feature data extracted from the second set of DWI images.