Fig. 3: The results of patch clustering and patch selecting. | Nature Communications

Fig. 3: The results of patch clustering and patch selecting.

From: Neuropathologist-level integrated classification of adult-type diffuse gliomas using deep learning from whole-slide pathological images

Fig. 3

a The silhouette plot (left) and the Calinski-Harabasz index plot (right) of the K-means clustering method with the cluster number ranging from 2 to 12. The silhouette coefficient, whose value ranges from -1 to 1, is used to assess the goodness of a clustering. A higher silhouette coefficient means better clustering. The Calinski-Harabasz index is calculated as the ratio of the between-cluster variance to the within-cluster variance. Similarly, a higher value of the Calinski-Harabasz index indicates better clustering performance. Well-grouped clusters are apart from each other and clearly distinguished. The silhouette coefficient and the Calinski-Harabasz index achieved their highest values of 0.447 and 4159.3, respectively, both at an optimal cluster number of nine. b Visualization of the nine clusters of the 43653 patches from 100 randomly selected patients in the training cohort. c Bar graph of patch-level classification accuracy of nine separate cluster-based classifiers. Three classifiers (shown by the red bar) trained on clusters 2,5,7 had higher accuracy than the benchmark classifier (shown by the green bar). Then, the patches within the three clusters 2,5,7 for each patient were selected for building the patient-level classifier. d The result of patch clustering and patch selection for three representative patients (top: A2; middle: O2; bottom: GBM). For each patient, the three images in the first row from left to right are the original whole-slide image, the distribution of the clustered patches (each color indicates a cluster), and the finally selected patches in the three clusters, respectively; the nine small images framed with different colors in the second row are representative patches from each of the nine clusters, where the finally selected three patches are shown in bold frames. Each experiment was repeated independently three times with the same results. Source data are provided as a Source Data file. Scale bars, 2 mm. The size of the patch is 1024 × 1024 pixels, with each pixel representing 0.50 microns.

Back to article page