Source-linked AI summary
Adapting Mask-RCNN for Automatic Nucleus Segmentation
Jeremiah W. Johnson
TL;DR
Microscopy nucleus segmentation is important, but methods developed for natural images must be evaluated on varied medical imagery. This paper adapts Mask-RCNN with feature-pyramid backbones and finds that it produces high-quality nucleus segmentations across widely varying microscopy images with little modification.
Problem
Automatic microscopy-image segmentation, including nucleus detection, is an important medical-image analysis task requiring evaluation of natural-image methods on varied microscopy data.
Method
The paper evaluates Mask-RCNN with feature pyramid network backbones, ROIAlign, a mask head, decoupled mask and class predictions, and limited preprocessing with augmentation.
Results
Mask-RCNN produced high-quality nucleus segmentations across widely varying microscopy images with very little modification.
Takeaways & Limitations
Mask-RCNN can be adapted to nucleus segmentation in varied microscopy images and may be investigated for related medical-image segmentation tasks.
Takeaways & Limitations
The dataset spans diverse organisms, imaging conditions, and biological contexts, creating a challenging lack of input standardization.
Abstract
from arXiv · showhide
Automatic segmentation of microscopy images is an important task in medical image processing and analysis. Nucleus detection is an important example of this task. Mask-RCNN is a recently proposed state-of-the-art algorithm for object detection, object localization, and object instance segmentation of natural images. In this paper we demonstrate that Mask-RCNN can be used to perform highly effective and efficient automatic segmentations of a wide range of microscopy images of cell nuclei, for a variety of cells acquired under a variety of conditions.
1. Introduction
The paper motivates adapting CNN-based instance-segmentation methods from natural images to microscopy nuclei, focusing on Mask-RCNN as a candidate for varied medical images.
- CNN-based models have driven major advances in natural-image detection, localization, semantic segmentation, and instance segmentation.
- U-Net was developed specifically for medical-image segmentation and achieved state-of-the-art results on neuronal-structure and cell-tracking challenges.
- Mask-RCNN extends Faster-RCNN for natural-image detection, localization, and instance segmentation, and outperformed existing single-model entries across every 2016 COCO Challenge task.
- 1.1. The Mask-RCNN Model.: Mask-RCNN uses region proposals, ROIAlign, a mask-producing network head, decoupled mask and class predictions, and a multitask loss.Its multitask loss is L = Lcls + Lbbox + Lmask.
- 1.1. The Mask-RCNN Model.: Because natural and medical images differ, the paper investigates whether Mask-RCNN can detect nuclei in microscopy images.
2. The Data
The experiments use a heterogeneous microscopy dataset of 729 annotated images spanning organisms, conditions, and biological contexts, with separate training-validation and test subsets.
- The BBBC038v1 dataset contains 729 microscopy images with expert annotations for each detected nucleus.The images were used in Stage 1 of the 2018 Data Science Bowl.
- The images span humans, mice, and flies, as well as tissues, embryos, cell division, and genotoxic stress.
- The dataset’s variation in organisms, imaging conditions, and contexts creates a challenge because CNNs generally perform best on uniform, standardized inputs.The relevant standardization dimensions include color, contrast, scale, and class balance.
- 664 images were used for training and validation, while 65 images were held out for testing.
3. Methodology
The study evaluates Mask-RCNN with feature-pyramid backbones, minimal preprocessing, and data augmentation, comparing ResNet-50-FPN with ResNet-101-FPN on validation data.
- The experiments use Mask-RCNN with a feature pyramid network backbone and compare ResNet-50-FPN against ResNet-101-FPN.The ResNet-50-FPN model has lower computational load, while ResNet-101-FPN improves results without other training changes.
- Images are upsampled twofold and channel means are normalized, while augmentation uses random crops, rotations, Gaussian blurring, and horizontal and vertical flips.Edge mirroring was tested but omitted after producing no improvement.
- 70.54% average mask IoU and 59.40% MSCOCO mean average precision were obtained with ResNet-101-FPN, versus 66.98% IoU and 56.06% AP with ResNet-50-FPN.
4. Conclusion & Future Work
The paper concludes that Mask-RCNN can segment nuclei in widely varying microscopy images with little modification and identifies related medical-segmentation applications for future study.
- Mask-RCNN produced high-quality nucleus segmentations across widely varying microscopy images with very little modification.
- The authors suggest investigating Mask-RCNN for left-ventricle, liver, and tumor segmentation in future work.They connect left-ventricle segmentation with estimating cardiac ejection fraction.