Abnormal Cervical Cell Detection Algorithm Based on Improved RetinaNet

Runkun Liu; Shijie Dang; Hongyuan Zhang; Yinyin Niu; Guanxun Mi; Sanhua Li; Zhenxin Chen; Lingxiao Zhao; Peng Li

doi:10.3788/CJL230718

Chinese Journal of Lasers, Volume. 50, Issue 15, 1507203(2023)

Abnormal Cervical Cell Detection Algorithm Based on Improved RetinaNet

Runkun Liu^1,2, Shijie Dang², Hongyuan Zhang², Yinyin Niu³, Guanxun Mi³, Sanhua Li³, Zhenxin Chen², Lingxiao Zhao^2、*, and Peng Li²

Author Affiliations

¹Division of Life Sciences and Medicine, School of Biomedical Engineering (Suzhou), University of Science and Technology of China, Hefei 230026, Anhui, China

²Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Science, Suzhou 215162, Jiangsu, China

³Henan Celnovte Biotechnology Co., Ltd., Zhengzhou 450001, Henan, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Structure of the AFF-RetinaNet

Download full size

Fig. 2. Residual block structure of ResNeSt-50

Download full size

Fig. 3. Diagram of split attention

Download full size

Fig. 4. Non-local attention block structure

Download full size

Fig. 5. CIoU Loss calculation diagram

Download full size

Fig. 6. Diagram of WSI inference process

Download full size

Fig. 7. Example of data annotation

Download full size

Fig. 8. Distribution density graph of label boxes size

Download full size

Fig. 9. Comparison of AFF-RetinaNet and RetinaNet detection results. (a) Ground truth; (b) RetinaNet; (c) AFF-RetinaNet

Download full size

Fig. 10. Comparison of WSI inference results and ground truth. (a) WSI inference results; (b) ground truth

Download full size

Fig. 11. Visualization examples of AFF-RetinaNet-based WSI inference results in ROI

Download full size

Table 1. Algorithm process of NMS

View table

Table 1. Algorithm process of NMS

NMS algorithm：remove redundant detection boxes

Input： detection boxes set B，IoU threshold T

Output： detection boxes set K after NMS

1：K←∅；

2：all boxes in B are descending ordered by confidence score；

3：whileB≠∅ do

4: b_ms←the first detection box in B；

5: add b_ms to K；

6: del b_msfrom B；

7: forbinBdo

8: if IoU of b and b_ms > Tdo

9: del b from B；

10: endif

11: endfor

12: endwhile

13: returnK

Table 2. Statistics of experiment data
View table
Table 2. Statistics of experiment data
Data type Number
WSI 233
ROI（total） 1103
Abnormal cell 11188
ROI for generating training set 883
Training set 8106
ROI for generating test set 220
Test set 2125

Table 3. Ablation experimental experiment
View table
Table 3. Ablation experimental experiment
Group ResNeSt-50 BFP CIoU Loss mAP /% mAP-s /% mAP-m /% mAP-l /% Recall %
1 × × × 80.2 13.6 38.5 49.1 59.0
2 √ × × 81.0 15.0 39.6 50.8 58.9
3 √ √ × 81.3 15.9 39.8 50.9 60.7
4 √ √ √ 83.4 24.4 41.5 51.8 60.2

Table 4. Performance comparison of mainstream algorithms on TCT dataset

View table

Table 4. Performance comparison of mainstream algorithms on TCT dataset

Algorithm	mAP / %	mAP-s / %	FPS	Parameter quantity /10⁶
YOLOv3	70.7	14.9	101.0	61.5
Cascade R-CNN	75.1	15.4	31.2	68.9
Faster R-CNN	77.4	14.3	37.3	41.1
LAD^［25］	77.6	14.9	21.6	31.8
FCOS	79.1	16.3	44.8	31.8
TOOD^［26］	79.9	15.5	20.7	31.8
RetinaNet	80.2	13.6	49.2	36.1
DDOD^［27］	80.7	14.3	39.6	31.9
AFF-RetinaNet	83.4	24.4	25.4	36.4

Table 5. Comparison of WSI inference results based on different algorithms
View table
Table 5. Comparison of WSI inference results based on different algorithms
Algorithm mAP /%
YOLOv3 56.5
LAD 59.0
FCOS 60.6
DDOD 62.0
TOOD 64.6
Cascade R-CNN 66.3
Faster R-CNN 67.4
RetinaNet 70.2
AFF-RetinaNet 70.8

Tools

Get Citation

Copy Citation Text

Runkun Liu, Shijie Dang, Hongyuan Zhang, Yinyin Niu, Guanxun Mi, Sanhua Li, Zhenxin Chen, Lingxiao Zhao, Peng Li. Abnormal Cervical Cell Detection Algorithm Based on Improved RetinaNet[J]. Chinese Journal of Lasers, 2023, 50(15): 1507203

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites