A deep-learning-based histopathology classifier for focal cortical dysplasia (FCD) unravels a complex scenario of comorbid FCD subtypes

Abstract

Objective: Recently, we developed a first artificial intelligence (AI)-based digital pathology classifier for focal cortical dysplasia (FCD) as defined by the ILAE classification. Herein, we tested the usefulness of the classifier in a retrospective histopathology workup scenario.

Methods: Eighty-six new cases with histopathologically confirmed FCD ILAE type Ia (FCDIa), FCDIIa, FCDIIb, mild malformation of cortical development with oligodendroglial hyperplasia in epilepsy (MOGHE), or mild malformations of cortical development were selected, 20 of which had confirmed gene mosaicism.

Results: The classifier always recognized the correct histopathology diagnosis in four or more 1000 × 1000-μm digital tiles in all cases. Furthermore, the final diagnosis overlapped with the largest batch of tiles assigned by the algorithm to one diagnostic entity in 80.2% of all cases. However, 86.2% of all cases revealed more than one diagnostic category. As an example, FCDIIb was identified in all of the 23 patients with histopathologically assigned FCDIIb, whereas the classifier correctly recognized FCDIIa tiles in 19 of these cases (83%), that is, dysmorphic neurons but no balloon cells. In contrast, the classifier misdiagnosed FCDIIb tiles in seven of 23 cases histopathologically assigned to FCDIIa (33%). This mandates a second look by the signing histopathologist to either confirm balloon cells or differentiate from reactive astrocytes. The algorithm also recognized coexisting architectural dysplasia, for example, vertically oriented microcolumns as in FCDIa, in 22% of cases classified as FCDII and in 62% of cases with MOGHE. Microscopic review confirmed microcolumns in the majority of tiles, suggesting that vertically oriented architectural abnormalities are more common than previously anticipated.

Significance: An AI-based diagnostic classifier will become a helpful tool in our future histopathology laboratory, in particular when large anatomical resections from epilepsy surgery require extensive resources. We also provide an open access web application allowing the histopathologist to virtually review digital tiles obtained from epilepsy surgery to corroborate their final diagnosis.

Keywords: brain; epilepsy; neuropathology; seizure.

FIGURE 1

Deep‐learning‐assisted diagnosis of major focal cortical dysplasia (FCD) subtypes from automatically extracted tiles. (A) This tile was extracted from a case histopathologically assigned to FCD type Ia (FCDIa). The classifier assigned this same diagnosis to the tile with a confidence score of .99 (hematoxylin and eosin [H&E] staining). (B) A modified class activation method (CAM) heatmap calculated from the same tile shown in panel A highlights the microcolumnar architecture as classifying parameter (color coding = the more reddish/darkish, the less important was the area for recognition). (C) The same CAM algorithm highlights many dysmorphic neurons in a tile of FCDIIa with a confidence score of .99 (see H&E staining of the same area in panel D). (E) This tile was classified as FCDIIb (confidence score = .99) and contains many balloon cells. (F) CAM recording highlights the same balloon cells as specifying feature (in yellowish/bright color). Scale bar in panel B = 150 μm; scale bars in panels C and F = 100 μm.

FIGURE 2

Gallery of tiles falsely assigned to focal cortical dysplasia type IIb (FCDIIb). (A) Eight‐year‐old female with frontal lobe epilepsy and histopathologically confirmed FCDIIa. This tile was extracted from a whole slide image obtained from a second surgery sample. Arrows point to reactive (gemistocytic) astrocytes in a region near the first resection border. (B) Class activation method (CAM) and contour calculation from the same tile highlight these astrocytes as classifying features (yellowish color) for the assignment to FCDIIb, that is, confusion with balloon cells. (C) Twelve‐year‐old male with frontal lobe epilepsy and intracerebral depth electrode recording from the suspected lesion. There were no dysmorphic neurons or balloon cells visible in this tile. (D) CAM and contour calculations highlight the area of confusion associated rather with tissue edema. (E, F) In the same patient shown in panel C, the algorithm confused dysmorphic neurons with balloon cells and assigned this tile also to FCDIIb. The inset in panel E reveals the neuronal nature of the cell highlighted by CAM and contour in the upper left of panel F. Scale bar in panel F = 150 μm and applies to all images.

FIGURE 3

Gallery of tiles falsely assigned to focal cortical dysplasia type IIa (FCDIIa). (A) Six‐year‐old female with clinically suspicious left temporo‐occipital FCDI. The histopathology diagnosis and classifier's MTCD (majority of tiles confirmed the diagnosis) was mild malformation of cortical development with oligodendroglial hyperplasia in epilepsy. (A) Class activation method (CAM)/contour labeling identified these objects as FCDIIa with a confidence score set to .9. The regular orientation of apical dendrites toward the pial surface visible on hematoxylin and eosin (H&E) staining shown in panel B classifies these contours as pyramidal cells of layer 5 (arrow). Perisurgical hypoxemic damage with condensed nuclei in a more reddish cytoplasm was another frequent feature in tiles falsely assigned to FCDIIa. (C, D) Fifty‐one‐year‐old male with magnetic resonance imaging‐negative temporal lobe epilepsy since age 15 years and the final histopathology diagnosis of heterotopic neurons in the white matter (mild malformations of cortical development). The classifier recognized FCDIIa in this tile of the neocortex (confidence score set to .9; CAM/contour in panel C), which was not confirmed in the H&E staining (D). Instead, an oblique cutting plane and hypoxemic damage (arrow) similar to that observed in panel A determined the classifier's decision. (E, F) Three‐year‐old male with left temporo‐occipital epilepsy and FCDIa. Similar to the images shown in panels B and D, the oblique cutting plane suggesting neuronal cell cluster and hypoxic damage governed the classifier's decision but can be ruled out by manual review. Scale bar in panel F = 150 μm and applies to all images.

FIGURE 4

Focal cortical dysplasia type Ia (FCDIa) coexisting with other major FCD subtypes. (A) One‐year‐old female with FCDIIa in the left temporo‐occipital region; 1.4% of tiles were recognized by the classifier as FCDIa due to a predominantly microcolumnar pattern. (B) Three‐year‐old male with FCDIIb in the right temporo‐occipital region; 3.1% of tiles were classified as FCDIa due to a predominantly microcolumnar pattern. (C) Two‐year‐old female with right frontal lobe epilepsy and histopathologically confirmed mild malformation of cortical development with oligodendroglial hyperplasia in epilepsy (SLC35A2 wild type); 12.9% of tiles were classified as FCDIa due to a predominantly microcolumnar pattern.

FIGURE 5

Regional proximity of coexisting focal cortical dysplasia (FCD) subtypes. FCD type IIb (FCDIIb), blue tiles; FCDIIa, purple tiles; FCDIa, green tiles. Note the regional proximity of FCDIa < FCDIIa < FCDIIb, the latter being assigned as final diagnosis in this case by MTCD (majority of tiles confirmed the diagnosis). Contour lines are used to better illustrate the detection limits of the underlying and interpolated raster graphics.