核心概念
提案された手法は、in-silico免疫組織化学(IHC)画像の生成において、既存の注釈を新しい領域に拡張することで、監督深層学習モデルのトレーニングに必要な大規模かつピクセル正確なデータセットのコストを著しく低減する可能性がある。
摘要
Abstract:
In-silico datasets can lower the cost of building large, precise datasets for deep learning models.
Introduction:
Training models with in-silico data minimizes costly dataset creation efforts.
CycleGANs and diffusion models are used for domain translation in computational pathology.
Methods:
ReStainGAN disentangles stain components in IHC images using an auxiliary IF domain.
Results:
ReStainGAN outperforms baseline methods in training nucleus segmentation models on created in-silico datasets.
Discussion:
ReStainGAN introduces a novel method for generating in-silico IHC images, showing superiority over baseline methods.
統計資料
F1 score: 0.848
Sensitivity: 0.840
Precision: 0.856