Meta multi-task nuclei segmentation with fewer training samples
Graphical abstract
Introduction
Computational pathology has been one of the hottest concepts in biological and clinical research in recent years. As the basic unit of the life system, cells/nuclei microenvironment delivers massive information. Many research works attempt to explore both explicit and implicit patterns of nuclei for disease diagnosis (Yamashita et al., 2020), prognostic prediction (Skrede, De Raedt, Kleppe, Hveem, Liestøl, Maddison, Askautrud, Pradhan, Nesheim, Albregtsen, et al., 2020, Lu, Koyuncu, Corredor, Prasanna, Leo, Wang, Janowczyk, Bera, Lewis, Velcheti, et al., 2020), tissue phenotyping (Javed et al., 2020) and etc. However, it is impractical to segment millions of nuclei manually due to the huge resolution of whole slide images. Therefore, it is valuable to construct a fully automatic model for nuclei segmentation.
Convolutional neural network has demonstrated the superiority in medical image segmentation (Moeskops, Wolterink, van der Velden, Gilhuijs, Leiner, Viergever, Išgum, 2016, Zhou, Siddiquee, Tajbakhsh, Liang, 2018, Chan, Hosseini, Rowsell, Plataniotis, Damaskinos, 2019). Meanwhile, Various CNN-based models (Graham, Vu, Raza, Azam, Tsang, Kwak, Rajpoot, 2019, Liu, Zhang, Song, Zhang, Zhang, O’Donnell, Cai, 2019) have been developed for nuclei segmentation and achieved outstanding performance under full supervision. However, most of the existing models are data-dependent. They have to optimize the neural networks for each individual dataset. When data from the different domains come, we still have to prepare a new dataset for the new training. Unfortunately, manual annotations are expensive, especially for histopathology images. It requires high expertise and is extremely tedious and time-consuming. Moreover, the neural network is often trained on top of a massive amount of training data with manual labels. Otherwise, it may drop into the overfitting problem. It is impractical to collect a large amount of annotated data for every single task. Therefore, it will be convenient and effective if fewer training samples can be used for model training while also greatly relieving the tremendous annotation workload of pathologists.
There have been a few attempts to develop neural network models with fewer training samples. Transfer learning (Pan and Yang, 2009) was proposed to address the lack of labeled data problem by transferring the data distribution from the source domain to the target domain. This kind of model fine-tuning strategy also relies on a certain degree of training data of the target domain. Recently, multi-task learning (Chen, Qi, Yu, Heng, 2016, Zhao, Chen, Li, Yu, Yao, Yan, Wang, Liu, Liang, Han, 2020) has been widely used for medical image segmentation by setting up auxiliary tasks from the same domain to improve the target task. This strategy can effectively increase model generalization. Meta-learning (Naik, Mammone, 1992, Hospedales, Antoniou, Micaelli, Storkey) aims to optimize the model itself instead of optimizing the final goal of the model. The learning to learn spirit is much closer to how humans learn the world.
In this paper, we aim to build a generalized model for nuclei segmentation with fewer training samples. Consequently, we absorb the spirits of meta-learning as well as multi-task learning and associate them together as a Meta Multi-Task Learning (Meta-MTL) scheme. To improve the model generalizability, we apply a model-agnostic meta-learning approach (Finn et al., 2017) as the outer optimization algorithm for the complete model. Then we construct a contour-aware multi-task learning model as the inner model for nuclei segmentation. There are two tasks in the inner model, a segmentation task and a contour detection task. These two parallel tasks can complement each other and let the neural network focus on more generalized features instead of task-specific ones. We construct an intermediate component with several feature fusion and interaction blocks (FFIB) to support feature interactions and feature sharing across two tasks. A feature attention strategy is introduced in the decoding phase to further amplify contour information in the segmentation task.
Extensive experiments are conducted to evaluate the effectiveness of our proposed model in three aspects, data dependency, domain adaptation and domain generalization. (1) For each dataset, we progressively deduct the among of training samples randomly with a 20% descending proportion. (2) Given a well-trained model on a specific dataset, we transfer this model to an unseen dataset with fewer training samples. (3) Given a model train on a specific dataset, we directly apply this model to an unseen dataset. Experimental results suggest that our model achieves comparable performances with state-of-the-art nuclei segmentation models with fewer training data. And our model demonstrates a better potential for domain adaptation and domain generalization. With such properties, our proposed model could be an initial segmentation model for the pathologist-in-the-loop labeling strategy with only limited start-up pixel-level labels, to save the annotation efforts. The contributions of this paper are three-fold:
- •
We propose a novel meta multi-task learning (Meta-MTL) scheme for nuclei segmentation with better domain adaptation and domain generalization abilities as well as less data dependency.
- •
In the contour-aware multi-task learning model, we propose a feature fusion and interaction block (FFIB) for features communication between two parallel tasks.
- •
Our proposed model achieves comparable performances with state-of-the-art models with fewer training samples.
Section snippets
Nuclei segmentation
Nuclei segmentation is a fundamental step in histopathology image analysis. Kumar et al. (2017) published a nuclei segmentation data set of histopathology images originating from multiple tissue sites, and proposed a simple CNN-based method to segment nuclei, background and cell boundaries, respectively. In addition, in order to improve the objectivity of the model evaluation, a new segmentation performance evaluation metric (Aggregated Jaccard Index, AJI) was proposed. This metric considered
Meta-MTL: meta multi-task nuclei segmentation
The huge resolutions and the expertise dependency of histopathology images make pixel-wise manual annotations extremely difficult. We want to fully discover the potential of the nuclei segmentation model, in order to minimize the dependency of the labeled data and improve model generalization, thereby finally alleviate the workloads of pathologists. To this end, we associate two very important techniques in meta-learning (MAML and multi-task learning) and propose a Meta-MTL model.
In this
Datasets and evaluation metrics
We evaluate our proposed method on two public nuclei segmentation datasets, Multi-Organ Nuclei Segmentation Dataset (MoNuSeg) (Kumar et al., 2017) and CPM-17 (Vu et al., 2019).
MoNuSeg contains 30 images from 7 tissue sites, including breast, kidney, liver, prostate, bladder, colon and stomach. The resolution of each image is . We split the whole dataset into a training set and a test set. The training set contains 16 images from 4 kinds of tissues, including the breast, kidney, liver,
Experiments
We evaluate the effectiveness of our proposed model on data dependency (Section 5.2), domain adaptation (Section 5.3) and domain generalization (Section 5.4) by a series of quantitative comparisons. The best performances of existing models are demonstrated in Section 5.1. Ablation studies (Section 5.5) are conducted to evaluation the superiority of the Meta-MTL design and the FFIB. Qualitative comparisons and results are demonstrated in each corresponding experiment. All the statistical results
Conclusion
In this paper, we propose a Meta-MTL model for nuclei segmentation. We aim to establish a general model with fewer training samples and less data dependency by associating an optimization-based meta-learning technique and a multi-task learning strategy. The proposed model achieves comparable performances with existing SOTA nuclei segmentation models with the complete training data. The superiority of our model appears when we gradually deduct the training samples. And it can still achieve
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the Key-Area Research and Development Program of Guangdong Province (No. 2021B0101420006), the National Science Fund for Distinguished Young Scholars (No. 81925023), the National Key R&D Program of China (No. 2021YFF1201003), Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application (No. 2022B1212010011), the National Natural Science Foundation of China (Nos. 82072090, 81771912, 62172120 and 82071892), the National
References (47)
- et al.
Hover-Net: simultaneous segmentation and classification of nuclei in multi-tissue histology images
Med. Image Anal.
(2019) - et al.
Sparse autoencoder for unsupervised nucleus detection and representation in histopathology images
Pattern Recognit.
(2019) - et al.
Cellular community detection for tissue phenotyping in colorectal cancer histology images
Med. Image Anal.
(2020) - et al.
Deep learning for prediction of colorectal cancer outcome: a discovery and validation study
Lancet
(2020) - et al.
Triple U-Net: hematoxylin-aware nuclei segmentation with progressive dense feature aggregation
Med. Image Anal.
(2020) - Behl, H. S., Najafi, M., Arnab, A., Torr, P. H., 2018. Meta learning deep visual words for fast video object...
- et al.
Meta-Seg: a generalized meta-learning framework for multi-class few-shot semantic segmentation
IEEE Access
(2019) Multitask learning
Mach. Learn.
(1997)- et al.
Weakly supervised multi-task learning for cell detection and segmentation
2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI)
(2020) - et al.
Histosegnet: semantic segmentation of histological tissue type in whole slide images
Proceedings of the IEEE International Conference on Computer Vision
(2019)
DCAN: deep contour-aware networks for accurate gland segmentation
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
A fast and accurate algorithm for nuclei instance segmentation in microscopy images
IEEE Access
Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification
Proceedings of the IEEE/CVF International Conference on Computer Vision
SAMS-NET: stain-aware multi-scale network for instance-based nuclei segmentation in histology images
2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)
Mask R-CNN
Proceedings of the IEEE International Conference on Computer Vision
Deep residual learning for image recognition
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
Densely connected convolutional networks
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
Panoptic segmentation
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
Siamese neural networks for one-shot image recognition
ICML Deep Learning Workshop, Lille
Cited by (20)
Transformer-based meta learning method for bearing fault identification under multiple small sample conditions
2024, Mechanical Systems and Signal ProcessingWBC YOLO-ViT: 2 Way - 2 stage white blood cell detection and classification with a combination of YOLOv5 and vision transformer
2024, Computers in Biology and MedicineLearning what and where to segment: A new perspective on medical image few-shot segmentation
2023, Medical Image Analysis
- 1
Contributed equally.