Attribution methodologies for neural networks designed for computer-aided diagnostic processes
Introduced here are diagnostic platforms able to attribute an output produced by a neural network to its input, as well as communicate the relationship between the output and input in a comprehensible manner. Neural networks are increasingly being used for critical tasks, such as detecting the presence/progression of medical conditions. Accordingly, the importance of explaining how these neural networks produce outputs has grown in importance. By explaining how outputs are produced by a neural network, a diagnostic platform can build trust with medical professionals responsible for interpreting the outputs, identify possible modes of neural network failure, and identify the latent variable(s) responsible for producing a given output.
1. A computer-implemented method for generating attribution visualizations with increased coverage, the method comprising:
acquiring an image generated during a diagnostic session;
generating a diagnosis for a condition by applying a diagnostic model to the image;
determining a contribution of each pixel in the image to the diagnosis by applying an attribution model to produce a respective attribution value for each pixel,
wherein each attribution value is representative of the importance of the corresponding pixel to the diagnosis;
clipping at least some of the attribution values having a largest positive magnitude so as to have a reduced positive magnitude;
after said clipping, transforming the attribution values into a color range, wherein a magnitude of a color value in the color range assigned to each pixel is proportional to the magnitude of the respective attribution value for such pixel; and
generating a visualization that visually distinguishes at least a subset of pixels in the image from other pixels in the image by visualizing at least the subset of pixels with their respective color values.
2. The computer-implemented method of claim 1 , further comprising identifying said subset of pixels in the image by:
detecting all pixels corresponding to attribution values that exceed a threshold.
3. The computer-implemented method of claim 1 , wherein
the attribution model is programmed to:
identify a reference image to serve as a baseline, and
estimate a contribution of each pixel in the image to the diagnosis based on a comparison to a corresponding pixel in the reference image; and
the method further comprises:
acquiring the reference image by continually applying a Gaussian blur to the image until blurriness reaches a predetermined threshold and then integrating along a Gaussian blur path.
4. A computer-implemented method comprising:
acquiring a prediction made by a neural network based on an image;
producing a series of attribution values by applying an attribution model to the prediction,
wherein the attribution model is programmed to generate an attribution value for each pixel in the image, and
wherein each attribution value is representative of the importance of the corresponding pixel to the prediction;
clipping at least some of the attribution values having a largest positive magnitude so as to have a reduced positive magnitude;
after said clipping, transforming the attribution values into a color range, wherein a magnitude of a color value in the color range assigned to each pixel is proportional to the magnitude of the respective attribution value for such pixel; and
generating a visualization that visually distinguishes at least a subset of pixels in the image from other pixels in the image by visualizing at least the subset of pixels with their respective color values.
5. The computer-implemented method of claim 4 , further comprising:
populating a matrix with the series of attribution values.
6. The computer-implemented method of claim 4 , wherein clipping at least some of the attribution values having a largest positive magnitude so as to have a reduced positive magnitude comprises:
sorting the series of attribution values by magnitude; and
filtering the series of attribution values so that all attribution values exceeding a threshold percentile are clipped.
7. The computer-implemented method of claim 4 , further comprising:
examining the series of attribution values to identify attribution values, if any, whose magnitude exceeds a first threshold; and
examining the series of attribution values to identify other attribution values, if any, whose magnitude falls beneath a second threshold.
8. The computer-implemented method of claim 7 , wherein pixels corresponding to the identified attribution values that exceed the first threshold are rendered in a first color, and wherein pixels corresponding to the other identified attribution values that fall beneath the second threshold are rendered in a second color different than the first color.
9. The computer-implemented method of claim 7 , wherein the other identified attribution values correspond to pixels in the image that oppose the prediction.
10. The computer-implemented method of claim 4 , further comprising:
causing display of the visualization on an interface by overlaying the image with the visualization.
11. The computer-implemented method of claim 10 , further comprising:
receiving input indicative of a request to remove the image from the interface; and
causing the image to be hidden from view.
12. The computer-implemented method of claim 4 , further comprising:
converting the image into grayscale format; and
causing display of the visualization on an interface by overlaying the converted image with the visualization.
13. The computer-implemented method of claim 4 , further comprising:
causing display of the image and the visualization in separate windows of an interface.
14. The computer-implemented method of claim 13 , wherein the separate windows share substantially similar markings to facilitate regional comparison between the image and the visualization.
15. A non-transitory computer-readable medium with instructions stored thereon that, when executed by a processor, cause the processor to perform operations comprising:
acquiring a prediction made by a neural network based on an image;
producing a series of attribution values by applying an attribution model programmed to:
establish a straight-line path from a reference image to the image,
compute a gradient at multiple points along the straight-line path, and
obtain an integrated gradients measure by cumulating the computed gradients;
clipping at least some of the attribution values having a largest positive magnitude so as to have a reduced positive magnitude; and
after said clipping, transforming the attribution values into a color range, wherein a magnitude of a color value in the color range assigned to each pixel is proportional to the magnitude of the respective attribution value for such pixel; and
generating a visualization that visually distinguishes at least a subset of pixels in the image from other pixels in the image by visualizing at least the subset of pixels with their respective color values.
16. The non-transitory computer-readable medium of claim 15 , wherein the operations further comprise:
selecting the reference image based on a characteristic of the input image, the diagnostic model, or any combination thereof.
17. The non-transitory computer-readable medium of claim 15 , wherein the integrated gradients measure is obtained by a for loop that makes repeated calls to a gradient operator.