Few-Shot Microscopy Image Cell Segmentation

Abstract

In microscopy image cell segmentation, it is common to train a deep neural network on source data, containing different types of microscopy images, and then fine-tune it using a support set comprising a few randomly selected and annotated training target images. In this paper, we argue that the random selection of unlabelled training target images to be annotated and included in the support set may not enable an effective fine-tuning process, so we propose a new approach to optimise this image selection process. Our approach involves a new scoring function to find informative unlabelled target images. In particular, we propose to measure the consistency in the model predictions on target images against specific data augmentations. However, we observe that the model trained with source datasets does not reliably evaluate consistency on target images. To alleviate this problem, we propose novel self-supervised pretext tasks to compute the scores of unlabelled target images. Finally, the top few images with the least consistency scores are added to the support set for oracle (i.e., expert) annotation and later used to fine-tune the model to the target images. In our evaluations that involve the segmentation of five different types of cell images, we demonstrate promising results on several target test sets compared to the random selection approach as well as other selection approaches, such as Shannon's entropy and Monte-Carlo dropout.

MetaMedSeg: Volumetric Meta-learning for Few-Shot Organ Segmentation

“…To compare with [18], we use the same hyperparameters as in the original paper for meta-training. It should be noted that, following the protocol from [18], we go sequentially over all the available source datasets rather than sampling them.…”

“…A task in meta-learning for segmentation can be defined in different ways. Our initial approach is based on [18]. A task is a set of k images and masks belonging to the same dataset.…”

“…In this work we focus on the meta-learning approach, which also comes in different flavours: metric learning (e.g., prototypical networks [5][6][7]), memory-based learning [8,9] and gradient-based methods. Few-shot learning for image segmentation has been an active research topic [10][11][12][13][14][15][16], but there are few works [17][18][19] focusing on few-shot medical image segmentation. In this work, we follow the meta-learning approach, relying on one of the recent modifications of the MAML (Model-Agnostic Meta-Learning) algorithm [20] -Reptile [21], which is a simple and yet effective algorithm that provides a wide field for experiments.…”

“…Reptile baseline, with and without our proposed volumetric task definition, 4. Few-shot cell segmentation by Dawoud et al [18], which is the closest related work, with and without our proposed volumetric task definition. Figure 1 shows an overview of our method.…”

Benchmarking of deep learning algorithms for 3D instance segmentation of confocal image datasets

“…In (Guerrero-Peña, Marrero-Fernández, Tsang, & Cunha, 2019) it is demonstrated how varying the contrasts of cell boundaries and a new loss function (weighted cross entropy) could be useful to obtain high accuracy segmentations when a 3D Unet model trained with a small and sparsely annotated training dataset. The issue of sparse annotations is also addressed in works like (Dawoud, Hornauer, Carneiro, & Belagiannis, 2020) and (Arbelle & Raviv, 2018) . c. Pipelines for segmenting images with densely packed cells/tissues : Cell instance segmentation in images where cells appear in dense clusters or in overlapping manner is a common research problem.…”