Deep convolutional encoder-decoder for prostate cancer detection and classification

1. A method for automated prostate tumor detection and classification in multi-parametric magnetic resonance imaging (MRI) images of a patient, comprising:

receiving a multi-parametric MRI image set of a patient, wherein the multi-parametric MRI image set includes a plurality of different types of MRI images of the patient;

performing simultaneous detection and classification of prostate tumors in the multi-parametric MRI image set of the patient using a trained multi-channel image-to-image convolutional encoder-decoder, wherein the trained multi-channel image-to image convolutional encoder-decoder includes multiple input channels to input multiple MRI images of the multi-parametric MRI image set of the patient and a plurality of output channels corresponding to a plurality of different tumor classes, and for each output channel, the trained multi-channel image-to image convolutional encoder-decoder generates a respective response map that provides detected locations of prostate tumors of the corresponding tumor class in the multi-parametric MRI image set of the patient, wherein

for the respective response map generated by the trained multi-channel image-to image convolutional encoder-decoder for each output channel is an image with intensity values that peak at each detected location of a prostate tumor of the corresponding tumor class and follow a Gaussian distribution in a vicinity of each detected location of a prostate tumor of the corresponding tumor class, and wherein

the plurality of different tumor classes includes a benign tumor class and a malignant tumor, and the trained multi-channel image-to image convolutional encoder-decoder generates a first response map that provides detected locations of benign prostate tumors in the multi-parametric MRI image set of the patient and a second response map that provides detected locations of malignant prostate tumors in the multi-parametric MRI image set of the patient.

2. The method of claim 1 , wherein the trained multi-channel image-to image convolutional encoder-decoder is trained based on a plurality of training multi-parametric MRI image sets and ground truth response maps for benign and malignant tumors that are generated for each of the plurality of training multi-parametric MRI image sets.

3. The method of claim 1 , further comprising: performing motion compensation on the plurality of different types of MRI images in the multi-parametric MRI image set of the patient prior to performing the simultaneous detection and classification of the prostate tumors in the multi-parametric MRI image set of the patient using the trained multi-channel image-to-image convolutional encoder-decoder.

4. The method of claim 3 , further comprising: extracting a region-of-interest (ROI) corresponding to the prostate and a surrounding area on each of a plurality of slices of the plurality of different types of MRI images in the multi-parametric MRI image set of the patient.

5. The method of claim 1 , wherein the multi-parametric MRI image set of the patient comprises a T2-weighted MRI image, an apparent diffusion coefficient (ADC) map derived from a diffusion weighted imaging (DWI) scan, a high b-value DWI image, and a K-Trans map generated from dynamic contrast enhanced (DCE) MRI scan, and the trained multi-channel image-to image convolutional encoder-decoder includes a respective input channel to input each of the T2-weighted MRI image, the ADC map, the high b-value DWI image, and the K-Trans map.

6. The method of claim 1 , wherein the multi-parametric MRI image set of the patient comprises a T2-weighted MRI image, an apparent diffusion coefficient (ADC) map derived from a diffusion weighted imaging (DWI) scan, and a high b-value DWI image, and the trained multi-channel image-to image convolutional encoder-decoder includes a respective input channel to input each of the T2-weighted MRI image, the ADC map, and the high b-value DWI image.

7. An apparatus for automated prostate tumor detection and classification in multi-parametric magnetic resonance imaging (MRI) images of a patient, comprising: means for receiving a multi-parametric MRI image set of a patient, wherein the multi-parametric MRI image set includes a plurality of different types of MRI images of the patient;

means for performing simultaneous detection and classification of prostate tumors in the multi-parametric MRI image set of the patient using a trained multi-channel image-to-image convolutional encoder-decoder, wherein the trained multi-channel image-to image convolutional encoder-decoder includes multiple input channels to input multiple MRI images of the multi-parametric MRI image set of the patient and a plurality of output channels corresponding to a plurality of different tumor classes, and for each output channel, the trained multi-channel image-to image convolutional encoder-decoder generates a respective response map that provides detected locations of prostate tumors of the corresponding tumor class in the multi-parametric MRI image set of the patient, wherein

for the respective response map generated by the trained multi-channel image-to image convolutional encoder-decoder for each output channel is an image with intensity values that peak at each detected location of a prostate tumor of the corresponding tumor class and follow a Gaussian distribution in a vicinity of each detected location of a prostate tumor of the corresponding tumor class, and wherein

the plurality of different tumor classes includes a benign tumor class and a malignant tumor, and the trained multi-channel image-to image convolutional encoder-decoder generates a first response map that provides detected locations of benign prostate tumors in the multi-parametric MRI image set of the patient and a second response ma that provides detected locations of malignant prostate tumors in the multi-parametric MRI image set of the patient.

8. The apparatus of claim 7 , further comprising: means for performing motion compensation on the plurality of different types of MRI images in the multi-parametric MRI image set of the patient prior to performing the simultaneous detection and classification of the prostate tumors in the multi-parametric MRI image set of the patient using the trained multi-channel image-to-image convolutional encoder-decoder.

9. The apparatus of claim 7 , wherein the multi-parametric MRI image set of the patient comprises a T2-weighted MRI image, an apparent diffusion coefficient (ADC) map derived from a diffusion weighted imaging (DWI) scan, a high b-value DWI image, and a K-Trans map generated from dynamic contrast enhanced (DCE) MRI scan, and the trained multi-channel image-to image convolutional encoder-decoder includes a respective input channel to input each of the T2-weighted MRI image, the ADC map, the high b-value DWI image, and the K-Trans map.

10. A non-transitory computer readable medium storing computer program instructions for automated prostate tumor detection and classification in multi-parametric magnetic resonance imaging (MRI) images of a patient, the computer program instructions when executed by processor cause the processor to perform operations comprising:

receiving a multi-parametric MRI image set of a patient, wherein the multi-parametric MRI image set includes a plurality of different types of MRI images of the patient;

performing simultaneous detection and classification of prostate tumors in the multi-parametric MRI image set of the patient using a trained multi-channel image-to-image convolutional encoder-decoder, wherein the trained multi-channel image-to image convolutional encoder-decoder includes multiple input channels to input multiple MRI images of the multi-parametric MRI image set of the patient and a plurality of output channels corresponding to a plurality of different tumor classes, and for each output channel, the trained multi-channel image-to image convolutional encoder-decoder generates a respective response map that provides detected locations of prostate tumors of the corresponding tumor class in the multi-parametric MRI image set of the patient, wherein

for the respective response map generated by the trained multi-channel image-to image convolutional encoder-decoder for each output channel is an image with intensity values that peak at each detected location of a prostate tumor of the corresponding tumor class and follow a Gaussian distribution in a vicinity of each detected location of a prostate tumor of the corresponding tumor class, and wherein

the plurality of different tumor classes includes a benign tumor class and a malignant tumor, and the trained multi-channel image-to image convolutional encoder-decoder generates a first response map that provides detected locations of benign prostate tumors in the multi-parametric MRI image set of the patient and a second response map that provides detected locations of malignant prostate tumors in the multi-parametric MRI image set of the patient.

11. The non-transitory computer readable medium of claim 10 , wherein the trained multi-channel image-to image convolutional encoder-decoder is trained based on a plurality of training multi-parametric MRI image sets and ground truth response maps for benign and malignant tumors that are generated for each of the plurality of training multi-parametric MRI image sets.

12. The non-transitory computer readable medium of claim 10 , wherein the operations further comprise: performing motion compensation on the plurality of different types of MRI images in the multi-parametric MRI image set of the patient prior to performing the simultaneous detection and classification of the prostate tumors in the multi-parametric MRI image set of the patient using the trained multi-channel image-to-image convolutional encoder-decoder.

13. The non-transitory computer readable medium of claim 10 , wherein the multi-parametric MRI image set of the patient comprises a T2-weighted MRI image, an apparent diffusion coefficient (ADC) map derived from a diffusion weighted imaging (DWI) scan, a high b-value DWI image, and a K-Trans map generated from dynamic contrast enhanced (DCE) MRI scan, and the trained multi-channel image-to image convolutional encoder-decoder includes a respective input channel to input each of the T2-weighted MRI image, the ADC map, the high b-value DWI image, and the K-Trans map.

14. The non-transitory computer readable medium of claim 13 , wherein the multi-parametric MRI image set of the patient comprises a T2-weighted MRI image, an apparent diffusion coefficient (ADC) map derived from a diffusion weighted imaging (DWI) scan, and a high b-value DWI image, and the trained multi-channel image-to image convolutional encoder-decoder includes a respective input channel to input each of the T2-weighted MRI image, the ADC map, and the high b-value DWI image.