應用於部分標註顯微鏡影像之半監督持續性學習實例分割器

Abstract

由於病理圖像的標註需要高度專業化，標註成本非常高，而且全標註病理數據集的規模通常遠小於自然圖像數據集。 在本文中，我們提出了一個基於Mask R-CNN的實例分割器。 為了減少標註的工作量，我們考慮了一種情境，即數據集只有部分標註，並提供了一些額外的未標註數據。在這種情況下，我們發現通過複製黏貼合成數據和BackErace增強，部分標記的數據可以更適當且有效地在訓練階段中被利用。此外，我們將半監督和持續性學習的蒸餾技術結合到一個框架中，使得我們的模型可以同時使用（部分）標記和未標記的數據來持續學習新的數據。我們在SegPC-2021病理學數據集上進行了實驗，結果顯示我們的方法能充分利用所有可用的數據，從而提高了性能。

Since the annotation of pathological images needs to be highly specialized, the annotation cost is very high, and the scale of fully annotated pathology datasets is usually much smaller than that of natural image datasets. In this study, we introduce a novel instance segmentation model, which is built upon the foundation of Mask R-CNN. To reduce the annotation effort, we consider a scenario where a dataset is only partially labeled with some extra unlabeled data provided. Under this scenario, we found that through copy-paste synthetic data and BackErace augmentation, partially labeled data can be more properly and effectively utilized in the training phase. In addition, we combine distillation techniques from semi-supervised and incremental learning into one framework, so that our model can continually learn new data using both (partially) labeled and unlabeled data simultaneously. We conducted experiments on SegPC-2021 pathology dataset, and the results demonstrate that our approach can sufficiently utilize all available data, leading to an improvement in performance.