Fig. 1: Deep learning pipeline.
From: Deep learning-based interpretable prediction of recurrence of diffuse large B-cell lymphoma
a H&E-stained core images were segmented and cut into non-overlapping patches of size 256 ×256 pixels. b Features were extracted from all patches of a core image using a ResNet50 encoder. The final feature vectors had a size of 1024. c CLAM model to classify the cores into ‘recurrence’ or ‘non-recurrence’. The model assignes attention scores to each patch of a core image, with blue frames indicating low attention patches and red frames indicating high attention patches. d Segmentation of the nuclei of the top 20 patches (based on the attention score) of each class using a pre-trained modified HoVer-Net architecture [24]. e We calculate morphological features from the segmented cell nuclei and statistically analyze them to find significant differences between recurred and non-recurred patients.
