醫學影像分割的弱監督持續學習

Abstract

在影像切割任務中，通常都需要人工去花費很大的心力標注出像素級別的標籤，來執行全監督式的學習訓練。在醫學影像中，由於這樣的標注任務通常需要專業人員來對影像進行標注，因此標注的成本也隨之提升。為了減緩在醫學影像標注的成本，在這篇論文中我們提出了基於點監督或塗鴉監督訓練的影像切割模型。在第一階段中我們藉助了在大量醫學影像上預訓練過的segment anything model(SAM) 來協助從點或塗鴉的標籤產生出像素級別的偽標籤，並在其中加入了VQ memory bank來儲存特徵與類別的關係，協助產生更高品質的偽標籤。在第二階段為了擺脫SAM在使用時需要輸入點或塗鴉標籤的依賴，另外訓練一個切割模型來基於偽標籤以及點或塗鴉標籤做學習。此外，為了因應訓練資料會隨時間增加，我們也加入了持續性學習的架構。這樣就能夠隨著低成本的標籤數量增加，持續更新切割模型。在BCSS乳癌切割資料集的實驗結果以及在PanNuke, MoNuSAC細胞切割資料集上的實驗結果都證明了我們提出的方法的效果，與其他基於塗鴉標籤以及點標籤的方法相比，都有所提升。

Image segmentation tasks require significant manual effort to label pixel-level annotations for fully supervised learning. In medical image segmentation, such annotation tasks usually need professional annotations, which increases the cost of labeling. To alleviate the burden of annotation in medical imaging, we introduce an image segmentation method based on point-supervised or scribble-supervised learning in this work. In the first phase, the proposed approach utilizes a pre-trained Segment Anything Model (SAM) on a large amount of medical imaging data to assist in generating pixel-level pseudo labels from point or scribble annotations. In addition, we incorporate a vector-quantizatoin memory bank to store the relationship between features and classes, aiding in the generation of pseudo labels. In the second phase, to eliminate the dependency on inputting point or scribble annotations when using SAM, we train a segmentation model based on the pseudo labels and point or scribble annotations. Furthermore, to accommodate the increasing amount of training data over time, we also incorporate a continual learning framework in this study. This allows the segmentation model to be continuously updated as more labels are added. Experimental results on the BCSS breast cancer segmentation dataset as well as the PanNuke and MoNuSAC nuclei segmentation datasets demonstrate the effectiveness of our method, showing improvements over other scribble-based and point-based approaches.