Black-box Domain Adaptative Cell Segmentation via Multi-source Distillation

Abstract

Cell segmentation plays a critical role in diagnosing various cancers. Although deep learning techniques have been widely investigated, the enormous types and diverse appearances of histopathological cells still pose significant challenges for clinical applications. Moreover, data protection policies in different clinical centers and hospitals limit the training of data-dependent deep models. In this paper, we present a novel framework for cross-tissue domain adaptative cell segmentation without access both source domain data and model parameters, namely Multi-source Black-box Domain Adaptation (MBDA). Given the target domain data, our framework can achieve the cell segmentation based on knowledge distillation, by only using the outputs of models trained on multiple source domain data. Considering the domain shift cross different pathological tissues, predictions from the source models may not be reliable, where the noise labels can limit the training of the target model. To address this issue, we propose two practical approaches for weighting knowledge from the multi-source model predictions and filtering out noisy predictions. First, we assign pixel-level weights to the outputs of source models to reduce uncertainty during knowledge distillation. Second, we design a pseudo-label cutout and selection strategy for these predictions to facilitate the knowledge distillation from local cells to global pathological images. Experimental results on four types of pathological tissues demonstrate that our proposed black-box domain adaptation approach can achieve comparable and even better performance in comparison with state-of-the-art white-box approaches. The code and dataset are released at: https://github.com/NeuronXJTU/MBDA-CellSeg.