CycleGAN for virtual stain transfer: Is seeing really believing?

Improving generalization capability of deep learning-based nuclei instance segmentation by non-deterministic train time and deterministic test time stain normalization

“…Although such approaches generally generate visually plausible image-to-image translation, they are still prone to create hallucinative image features, and they are quite sensitive to the model architecture and training procedure. Thus, their application for some tasks, including specific image segmentation in histological images, has shown to be limited [16] , [47] .…”

“…However, some studies have faced constraints due to the use of limited test sets or a restricted number of tissues/organs for assessing generalization capabilities [5] , [13] , [24] , [26] , [46] . Additionally, certain investigations have simultaneously addressed multiple tasks, such as classification and segmentation [25] , while others have focused on proposing solutions for distinct tasks, such as virtual stain transfer in histological images [46] , [47] , [55] .…”

CohortFinder: an open-source tool for data-driven partitioning of digital pathology and imaging cohorts to yield robust machine-learning models

“…ComBat 7 has been used to reduce the variability of radiomic features by considering different scan parameters as separate ‘batches’ and applying a non-parametric normalization between them; which may unfortunately introduce unintended correlations leading to higher false positive rates (FPR) 14 . GAN 3 , 13 have been used to generate synthetic medical images for both CT liver images 13 and hematoxylin and eosin (H&E) pathology images 3 ; however, there is a risk of GANs “hallucinating”, i.e., generating unrealistic, or untrustworthy images that are not representative of the appearance of real disease biology 15 . Another set of techniques for managing BE focuses on data augmentation 14 , 16 ; a suite of techniques that involves increasing the size of training cohorts through the generation of additional synthetic samples (e.g., creating variants of an image based on permutations of brightness or contrast levels of real samples).…”