# ACDS - Automatic Cell Detection and Segmentation

Abstract
Paper
Dataset
Code
Results
Quantitative Comparison
Qualitative Comparison
Acknowledgments

Our proposed method to detect and segment the phytoplankton cells from microscopic images of non-setae species.

## Abstract

Saliency-based marker-controlled watershed method was proposed to detect and segment phytoplankton cells from microscopic images of non-setae species. This method first improved IG saliency detection method by combining saturation feature with color and luminance feature to detect cells from microscopic images uniformly and then produced effective internal and external markers by removing various specific noises in microscopic images for efficient performance of watershed segmentation automatically. We built the first benchmark dataset for cell detection and segmentation, including 240 microscopic images across multiple phytoplankton species with pixel-wise cell regions labeled by a taxonomist, to evaluate our method. We compared our cell detection method with seven popular saliency detection methods and our cell segmentation method with six commonly used segmentation methods. The quantitative comparison validates that our method performs better on cell detection in terms of robustness and uniformity and cell segmentation in terms of accuracy and completeness. The qualitative results show that our improved saliency detection method can detect and highlight all cells, and the following marker selection scheme can remove the corner noise caused by illumination, the small noise caused by specks, and debris, as well as deal with blurred edges.

## Paper

• Haiyong Zheng*, Nan Wang, Zhibin Yu, Zhaorui Gu, Bing Zheng. Robust and automatic cell detection and segmentation from microscopic images of non-setae phytoplankton species. IET Image Processing, 2017, DOI: 10.1049/iet-ipr.2017.0127. (SCI)

## Dataset

We construct a new dataset that contains 240 phytoplankton microscopic images with 225 of single cell while 15 of multiple cells and human labeled ground-truth cell regions. These images are acquired and selected with the help of phytoplankton experts in different sizes from $256\times 256$ to $4080\times 3072$ and different species of non-setae phytoplankton. The pixel-wise ground truth masks are produced guided by phytoplankton taxonomist to contain the biomorphic characteristics of cells. These data provide useful resource to study the phytoplankton for automatic detection and segmentation. Some sample images are shown in the above figure. The data are available for downloading here.

## Results

### Quantitative Comparison

• Quantitative comparison of cell detection by different methods on our proposed dataset. (a) 225 microscopic images with single cells. The yellow $\bigstar$ representing the actual threshold generated by our method for binarization shows the nearly best segmentation on these PR curves. (b) 15 microscopic images with multiple cells.

• Quantitative comparison of cell segmentation by different methods on our proposed dataset. (a) 225 microscopic images with single cells. (b) 15 microscopic images with multiple cells.

• Shape matching using modified Hausdorff distance (MHD). For 225 microscopic images with a single cell (see S1 Table), the contour (shape)-matching number most similar to the ground truth is 12 by Canny, 2 by Ours1/Ours2, and 211 by our final segmentation (Ours3). For 15 microscopic images with multiple cells (see S2 Table), the corresponding matching number is 2 by Ours2 and 13 by Ours3, indicating the better performance of our proposed method on shape similarity of phytoplankton cells.
• Cell counting comparison on 15 microscopic images with multiple cells. The numbers in bold font indicate the best results.
No. Canny ITS Otsu Sauvola MET Kmeans Ours1 Ours2 Ours3 GT
#1 10 55 11 51 12 12 16 14 12 12
#2 8 7 7 57 7 8 10 10 9 9
#3 158 839 745 203 177 455 267 198 125 141
#4 207 3390 12 11 46 9 9 5 3 3
#5 12 5323 6 46292 44 14232 8 2 2 2
#6 35 107 6 5220 49 3014 17 2 2 2
#7 10 228 10 14 27 10 11 9 5 5
#8 8 11 5 4 11 3 21 3 2 2
#9 35 345 338 53 32 34 39 35 31 31
#10 6 614 9 7 9 4 3 2 2 2
#11 18 627 336 90 58 114 32 14 12 11
#12 35 224 37 42 121 33 30 29 29 31
#13 8 390 8 10 24 7 13 10 5 5
#14 11 11 11 26 11 11 12 11 10 11
#15 12 6934 4 41825 114 15283 47 21 5 2

### Qualitative Comparison

• Experimental results of our proposed method for single cell detection and segmentation. The first column shows the original RGB microscopic images of the following non-setae species in each row: (a) Dictyocha fibula. (b) Chattonella marina. Columns 2 to 6 present the image results of salient objects detected by the IG method, the salient objects detected by saturation, our combined salient objects, the binarization of combined salient objects, and the markers containing internal (black regions in the objects) and external (black regions outside the objects) markers imposed on the gray-level microscopic images, respectively. The last column shows the final segmentation results of our proposed method.

• Visual comparison with the commonly used segmentation methods on single cell segmentation. The first column shows the original RGB microscopic images of the following non-setae species in each row: (a) Dictyocha fibula. (b) Chattonella marina. (c) Ceratium tripos. (d) Scrippsiella trochoidea. For comparison, the remaining columns present the results obtained by the following segmentation methods consecutively: Canny, ITS, Otsu, Sauvola, MET, K-means, and our proposed method.

• Experimental results of our proposed method for multiple cell detection and segmentation. The first column shows the original RGB microscopic images of the following non-setae species in each row: (a) Prorocentrum triestinum. (b) Amphidinium carterae. (c)(d) Chattonella marina. Columns 2 to 6 present the image results of the salient objects detected by the IG method, the salient objects detected by saturation, our combined salient objects, the binarization of combined salient objects, and the markers containing internal (black regions on the objects) and external (black lines between the objects) markers imposed on the gray-level microscopic images, respectively. The last column shows the final segmentation results of our proposed method.

## Acknowledgments

We wish to thank the Algal Collection of Research Center for Harmful Algae and Aquatic Environment in Jinan University and Key Laboratory of Marine Environment and Ecology, Ministry of Education in Ocean University of China for providing the samples of phytoplankton species and the instruments to observe and acquire the corresponding microscopic images. This work was supported by the National Natural Science Foundation of China under grant numbers 61301240 and 61271406 and China Postdoctoral Science Foundation under grant number 2016M590658.