NuSegDA: Domain adaptation for nuclei segmentation.

Area of Science:

• Medical image analysis
• Computational pathology
• Artificial intelligence in oncology

Background:

• Accurate nuclei segmentation is vital for cancer diagnosis and treatment planning.
• Supervised learning for nuclei segmentation requires large, annotated datasets, which are scarce and costly to produce.
• Unlabeled datasets present a significant challenge for training effective nuclei segmentation models.

Purpose of the Study:

• To develop a nuclei segmentation model capable of training on unlabeled datasets using unsupervised domain adaptation (UDA).
• To adapt segmentation models from a labeled source domain to an unlabeled target domain, even with differing data characteristics.
• To improve the feasibility of nuclei segmentation in clinical settings where annotated data is limited.

Main Methods:

• Proposed NuSegUDA model employing Unsupervised Domain Adaptation (UDA) across feature and output spaces.
• Utilized a reconstruction network with adversarial learning for accurate source-to-target domain image translation.
• Incorporated novel image reconstruction adversarial loss and target-translated source supervised loss.

Main Results:

• NuSegUDA demonstrated significant performance improvements over baseline methods on public nuclei segmentation datasets.
• Experimental validation confirmed the effectiveness of the proposed adversarial loss and translation-based supervised loss.
• An extension, NuSegSSDA, was developed for scenarios with limited target domain annotations, showing promise in semi-supervised settings.

Conclusions:

• Unsupervised Domain Adaptation is a viable strategy for nuclei segmentation with unlabeled data.
• The NuSegUDA model effectively bridges the domain gap between labeled source and unlabeled target datasets.
• The proposed method offers a practical solution for leveraging large unlabeled datasets in digital pathology.