(A-12) Ensemble U-Net Architecture for Nuclei Segmentation with Novel dataset

Introduction:: UNet is a common method for nuclear segmentation. This network is trained on a set of substantial pre-processed and human-labeled data. The labelling is laborious where each pixel is annotated as either part of a nucleus or background. Such supervised learning reprocessing is time-consuming and may introduce biases. During training, Unet learns to predict the correct label for each pixel based on the input image. Once trained, the network can be used to segment new images by applying them to each pixel of the image and generating a segmentation mask that indicates which pixels belong to nuclei. Given the dearth of annotated data, which makes it difficult to train and deploy the network on resource-limited devices, we propose to improve cell segmentation by using self-supervised learning. We propose to develop an SSL-assisted UNet framework to improve the segmentation of nuclei, DNA damage within the nuclei, and genomic instability.

Materials and Methods:: Materials and Methods: In UNet, the input image is processed through a series of convolutional and pooling layers to extract high-level features. The resulting feature map is then passed through a decoder network that samples the feature map to the same resolution as the original input image, while also merging in the feature maps from the corresponding encoder layers. SSL will be used to pre-train the encoder of UNet in an unsupervised manner to learn data representation. SSL-based UNet allows one to learn both high-level and low-level features, making it well-suited for image segmentation tasks. We aim to develop SSL-based UNet segmentation via self-supervised representation learning in fluorescent microscopy images. The objective of segmentation entails categorizing every pixel in an image; hence the pre-training network must learn representations of features at the pixel level. Our approaches include identifying organoids, disassociating contacting cells, and segmenting such subcellular compartments as nuclei, DNA damage, and genomic instability (Figure 1A). Our system combines UNet with an SSL-based deep learning approach to learn and predict the positions of nuclei and functional molecular markers. We trained and validated our approach using different sets of images of 3D datasets and 2D image slices. Those images contain stained nuclei and molecular markers, which were imaged using Zeiss 710 equipped with a 40X objective.

Results, Conclusions, and Discussions:: Results and Discussion: We investigate various data augmentations for self-supervised semantic segmentation that haven’t been used before. The final result is a pre-training method that uses both the tested data augmentations, e.g., ground truth images (Figure 1B), and the local propagation module. As a result, this approach eliminates the potential for variability and partiality among end-users. The SSL-based UNet model is fine-tuned for the subsequent image segmentation. The pipeline of the model uses annotated data samples after pre-training the encoder network with feature representations in an unsupervised manner (Figure 1A). With exhaustive experimental trials and correlation metrics demonstrated, it is evident that self-supervised models like Barlow's Twin, SimCLR, and SwAV tend to improve the segmentation performance of U-Net models (Figure 1C). Barlow's Twin model demonstrates dice coefficient increased by 15% and precision increased by 11% compared to the traditional UNet model.

The suggested model makes use of sophisticated segmentation techniques that can be self-trained with a modest dataset and limited computation resources. The investigation in this work would be based on how unlabeled data can be used to pre-train a network to learn pixel-level feature representations suitable for segmentation. We expect the automated nuclei segmentation models to become competitive with other state-of-the-art semantic segmentation methods with limited data size and reduce the cost of computation resources.

Acknowledgements (Optional): : We gratefully thank College of Engineering & Applied Science at University of Wisconsin Milwaukee for providing PI's startup grant.

References (Optional): :