Diagnosis of Cervical Cancer based on Ensemble Deep Learning Network using Colposcopy Images

Table 1

Summary of the related works for screening cervical cancer.

S.no	Methods	Dataset	Advantages	Disadvantages
1	Inception V3 model [1]	Herlev dataset	(i) High accuracy; (ii) Good universality; Low complexity	(i) The deep network needs further study to investigate cervical cells.
2	Transfer learning, pretrained DenseNet [2]	Fujian Maternal and child health hospital Kaggle	(i) More feasibility and effective	(i) Limited data
3	CNN-extreme learning machine- (ELM-) based system [6]	Herlev dataset	(i) Fast learning; (ii) Easy convergence; (iii) Less randomized	(i) More complexity; (ii) Need more investigation
4	Gene-assistance module, voting strategy [7]	Chinese hospital and Universitario De Caracas, Venezuela	(i) More scalable and practical	(i) Limited datasets
5	Random forest and Adaboost [14]	Radiotherapy dataset	(i) Better treatment planning	(i) Need to extract features; (ii) Painful treatment
6	ColpoNet [16]	Colposcopy images	(i) Better accuracy; (ii) Efficient classification	(i) Need to improve accuracy by extracting relevant information
7	CNN Model [17]	Papanicolaou-stained cervical smear dataset	(i) Better sensitivity and specificity	(i) Reported 1.8% false-negative images
8	Fourier transform and machine learning methods. [18]	Microscopic images	(i) Fully automatic system; (ii) Saving precious time for the microscopist	(i) The level of complexity is more
9	CNN-SVM model [21]	Herlev and one private dataset	(i) Good robustness; (ii) Highest accuracy	(i) Need improvement to adjust parameter; (ii) Need of hand-crafted features
10	Stacked Autoencoder [27]	UCI database	(i) High accuracy; (ii) Reduced data dimension	(i) Training time is very high due to reducing the dimension
11	PSO with KNN algorithm [33]	Cervical smear images	(i) Better accuracy; (ii) Good feature selection	(i) Time-consuming due to two-phase feature selection
12	Ensemble model [34]	PAP smear image	(i) For 2 class problem achieves the accuracy of 96%; (ii) For 7 class problem achieves an accuracy of 78%	(i) Overall of cells are difficult to identify
13	Multimodal deep network [37]	National Cancer Institute	(i) Good correlation; (ii) High accuracy; (iii) Learn better complementary features	(i) More complexity in image fusion