Sequential integration of adversarial networks with handcrafted features (SANwicH): identifying sites of prognostic significance for predicting cancer recurrence
Embodiments discussed herein facilitate generation of a prognosis for a medical condition based on determination of one or more histomorphometric features for tiles of a whole slide image (WSI) that have been identified as the most prognostically significant tiles of the WSI. A first set of embodiments discussed herein relates to training of a fully convolutional network (FCN) to determine the prognostic significance of pixels of a WSI. A second set of embodiments discussed herein relates to determination of a prognosis based on analysis of regions identified as the most prognostically significant by a trained FCN.
1. A non-transitory computer-readable medium storing computer-executable instructions that, when executed, cause a processor to perform operations, comprising:
accessing a digitized whole slide image (WSI) of tissue associated with a medical condition, wherein the digitized WSI comprises a plurality of pixels, wherein each pixel of the plurality of pixels has an associated intensity;
analyzing the digitized WSI with an adversarially trained fully convolutional network (FCN) to determine, for each pixel of the plurality of pixels, an associated prognostic significance (PS) for that pixel;
generating a PS map corresponding to the digitized WSI that indicates, for each pixel of the plurality of pixels, the associated PS for that pixel;
dividing the digitized WSI into a plurality of tiles, wherein each tile of the plurality of tiles is a square comprising an associated k×k pixels of the plurality of pixels, wherein each tile has a tile PS based on the associated PS of each pixel of the associated k×k pixels of that tile;
selecting N tiles of the plurality of tiles, wherein the selected N tiles have N highest tile PS among tiles of the plurality of tiles;
computing, for each of the N tiles, an associated prognostic metric for that tile based on one or more histomorphometric features; and
generating a prognosis for the digitized WSI in connection with the medical condition based on the associated prognostic metric computed for each tile of the N tiles.
2. The non-transitory computer-readable medium of claim 1 , wherein each histomorphometric feature of the one or more histomorphometric features comprise one or more handcrafted features selected for prognostic significance with respect to the medical condition via at least one of a minimum redundancy maximum relevance algorithm or quadratic discriminant analysis (QDA).
3. The non-transitory computer-readable medium of claim 1 , wherein each histomorphometric feature of the one or more histomorphometric features is one of a global graph feature, a nuclear cluster graph feature, a nuclei shape feature, a nuclei orientation entropy feature, or a nuclei texture feature.
4. The non-transitory computer-readable medium of claim 1 , wherein the medical condition is early stage non-small cell lung cancer (NSCLC).
5. The non-transitory computer-readable medium of claim 1 , wherein N is less than or equal to 20.
6. The non-transitory computer-readable medium of claim 5 , wherein N is 10.
7. The non-transitory computer-readable medium of claim 1 , wherein k is 1000.
8. The non-transitory computer-readable medium of claim 1 , wherein, for each pixel, the associated PS for that pixel is generated by the FCN via forward inference.
9. The non-transitory computer-readable medium of claim 1 , wherein, for each pixel, the associated PS for that pixel is generated by the FCN based on a binary prediction for an input image patch around that pixel.
10. The non-transitory computer-readable medium of claim 1 , wherein the prognosis is one of recurrence or non-recurrence.
11. An apparatus that facilitates generation of a prognosis in connection with a medical condition, the apparatus comprising:
a processor;
a memory configured to store a digitized whole slide image (WSI) of tissue associated with the medical condition, wherein the digitized WSI comprises a plurality of pixels, wherein each pixel of the plurality of pixels has an associated intensity;
an input/output (I/O) interface;
a set of circuits; and
an interface that connects the processor, the memory, the I/O interface, and the set of circuits, the set of circuits comprising:
an image acquisition circuit configured to access the digitized WSI;
an adversarial fully convolutional network circuit configured to:
analyze the digitized WSI to determine, for each pixel of the plurality of pixels, an associated prognostic significance (PS) for that pixel;
generate a PS map corresponding to the digitized WSI that indicates, for each pixel of the plurality of pixels, the associated PS for that pixel;
divide the digitized WSI into a plurality of tiles, wherein each tile of the plurality of tiles is a square comprising an associated k×k pixels of the plurality of pixels, wherein each tile has a tile PS based on the associated PS of each pixel of the associated k×k pixels of that tile; and
select N tiles of the plurality of tiles, wherein the selected N tiles have N highest tile PS among tiles of the plurality of tiles;
a histomorphometric feature calculation circuit configured to:
compute, for each of the N tiles, an associated prognostic metric for that tile based on one or more histomorphometric features; and
generate the prognosis for the digitized WSI in connection with the medical condition based on the associated prognostic metric computed for each tile of the N tiles; and
a display circuit configured to output the prognosis for the digitized WSI in connection with the medical condition.
12. The apparatus of claim 11 , wherein each histomorphometric feature of the one or more histomorphometric features comprise one or more handcrafted features selected for prognostic significance with respect to the medical condition via at least one of a minimum redundancy maximum relevance algorithm or quadratic discriminant analysis (QDA).
13. The apparatus of claim 11 , wherein each histomorphometric feature of the one or more histomorphometric features is one of a global graph feature, a nuclear cluster graph feature, a nuclei shape feature, a nuclei orientation entropy feature, or a nuclei texture feature.
14. The apparatus of claim 11 , wherein the medical condition is early stage non-small cell lung cancer (NSCLC).
15. The apparatus of claim 11 , wherein N is less than or equal to 20.
16. The apparatus of claim 15 , wherein N is 10.
17. The apparatus of claim 11 , wherein, for each pixel, the associated PS for that pixel is generated by an adversarially trained fully convolutional network (FCN) via forward inference.
18. The apparatus of claim 11 , wherein, for each pixel, the associated PS for that pixel is generated by an adversarially trained fully convolutional network (FCN) based on a binary prediction for an input image patch around that pixel.