基於類別無關小樣本學習之數位病理影像實例分割

Abstract

精準地分割出單一細胞或細胞核對於數位病理影像分析是至關重要的環節。近年來歸功於深度神經網路的進步，數位病理影像實例分割（instance segmentation）已取得大幅度的進展。然而，傳統的深度學習網路無法在從沒見過的新染色型態影像上順利地辨認物體。再者，對於醫療或生物科學領域，蒐集巨量的標註資料往往耗時且費工。為了解決這些挑戰，我們提出一個基於類別無關（class-agnostic）小樣本學習（few-shot learning）的方法來執行數位病理影像的實例分割任務：給定一個屬於任意一種染色型態的目標（target）影像，只需額外提供一個或少量的參考（reference）影像作為引導，這些參考影像具有與目標影像同樣染色型態的細胞或細胞核，我們的模型就能夠透過類別無關（class-agnostic）的方式單獨地分割出目標影像中相對應的細胞或細胞核。我們提出的模型利用在多尺度別（multi-scale）的共注意（co-attention）機制以及 class-agnostic relation R-CNN 來取得目標影像與參考影像的多尺度相關特徵。另外，我們也使用可變形卷積（deformable convolution）以捕捉細胞排列的幾何結構。我們更廣泛地蒐羅 11 個公開的數位病理影像資料集以建構出新的小樣本實例分割資料集。透過實驗的結果也顯示出我們的小樣本學習方法勝過先前的方法而達到當前最優的結果。此外，我們的方法在性能表現上也拉近只需部分資料訓練的小樣本學習模型與使用全部的資料訓練的傳統深度學習模型之間的差距。這些令人信服的結果揭示了我們的方法的優勢，該方法只需提供少量標註的樣本便能有效地對於數位病理影像進行實例分割，大幅減輕人為標註的困難。

Accurately segmenting individual cells or nuclei is a crucial step in the digital pathological image analysis procedure. In recent years, it has yielded substantial progress attributed to the advances in deep neural networks. However, traditional deep learning networks cannot smoothly recognize the staining modalities never seen. Besides, in the medical or bio-scientific domain, collecting tremendous annotated data is time-consuming and labor-intensive. To address the challenges, we present a class-agnostic few-shot learning for the digital pathological image instance segmentation: given a target image from any staining modality, our model can individually segment the cells or nuclei within target in a class-agnostic manner by providing one or very few reference images as guidance. We design a multi-scale co-attention mechanism and a class-agnostic relation R-CNN to leverage the correlated multi-scale features and utilize the deformable convolution to capture the geometric structures. Moreover, we establish a new few-shot instance segmentation dataset by widely gathering 11 public digital pathological image datasets. The experimental results demonstrate the superiority of our method in comparison with the state-of-the-art. Furthermore, our approach closes the performance gap between few-shot learning model trained with partial data and the Oracle trained with all fully labelled data. Such convincing evidences reveal the advantage of our approach that can conduct the cell instance segmentation with very few annotations to mitigate the difficulty of the manual annotation.