Evaluation of Deep Learning Architectures for Complex Immunofluorescence Nuclear Image Segmentation.

Florian Kromp,Wolfgang Dorr,Inge M Ambros,Klaus Beiske,Eva Bozsaky,Allan Hanbury,Sabine Taschner-Mandl,Lukas Fischer,Peter F Ambros

doi:10.1109/tmi.2021.3069558

Abstract

Separating and labeling each nuclear instance (instance-aware segmentation) is the key challenge in nuclear image segmentation. Deep Convolutional Neural Networks have been demonstrated to solve nuclear image segmentation tasks across different imaging modalities, but a systematic comparison on complex immunofluorescence images has not been performed. Deep learning based segmentation requires annotated datasets for training, but annotated fluorescence nuclear image datasets are rare and of limited size and complexity. In this work, we evaluate and compare the segmentation effectiveness of multiple deep learning architectures (U-Net, U-Net ResNet, Cellpose, Mask R-CNN, KG instance segmentation) and two conventional algorithms (Iterative h-min based watershed, Attributed relational graphs) on complex fluorescence nuclear images of various types. We propose and evaluate a novel strategy to create artificial images to extend the training set. Results show that instance-aware segmentation architectures and Cellpose outperform the U-Net architectures and conventional methods on complex images in terms of F1 scores, while the U-Net architectures achieve overall higher mean Dice scores. Training with additional artificially generated images improves recall and F1 scores for complex images, thereby leading to top F1 scores for three out of five sample preparation types. Mask R-CNN trained on artificial images achieves the overall highest F1 score on complex images of similar conditions to the training set images while Cellpose achieves the overall highest F1 score on complex images of new imaging conditions. We provide quantitative results demonstrating that images annotated by under-graduates are sufficient for training instance-aware segmentation architectures to efficiently segment complex fluorescence nuclear images.

Highlights

M ICROSCOPY has become a powerful tool to gain insights into cellular or sub-cellular structures by visualizing cellular compartments such as the nucleus, the cytoplasm, sub-cellular appearance of proteins or DNA elements [1]
We compared and evaluated the segmentation effectiveness of five deep learning architectures and two conventional methods on an expert-annotated nuclear image dataset composed of images from multiple sample preparation types showing a wide range of variations
We investigated the concept of image complexity to evaluate state-of-the-art deep learning architectures and conventional methods with respect to the segmentation challenge these images present

Summary

Introduction

M ICROSCOPY has become a powerful tool to gain insights into cellular or sub-cellular structures by visualizing cellular compartments such as the nucleus, the cytoplasm, sub-cellular appearance of proteins or DNA elements [1]. By applying automated microscopes and image analysis workflows, quantitative results can be generated at the single cell level. While pathology departments routinely use Hematoxylin and Eosin (H&E) histological or immunohistochemical (IHC) stainings, research laboratories mainly rely on immunofluorescence (IF) stainings This is because up to 90 or more (sub-)cellular compartments can be visualized simultaneously using multiplex-IF staining techniques and epifluorescence microscopy. Quantitative, microscopy based image analysis workflows generally consist of the following steps: sample preparation, microscopy image acquisition, nuclear and/or cytoplasmic image segmentation, feature extraction and cell population analysis. Each step within such a workflow can impact quantification and interpretation of experiments [4]. To generate quantitative results at the single cell level, segmentation algorithms must segment each nucleus instance

Objectives

Results

Conclusion