TABLE 3.

CTC enumeration approach comparisons.

Literatures	Approach overview	Data allocations	CTC annotation methods	Performance
G. Da Col, et al. [5]	Using traditional machine learning with man-crafted image features	A total of 2,598 cells from 45 MBC patients with 846 CD45 pos cells and 344 eCTC in training.	interactive mode to select interested cells and calculated features	91% accuracy with eCTC	84% with CD45pos
Z. Guo, et al. [11]	Automatic enumeration of CTCs with pure deep learning (RESNET18) and transfer learning assistance	694 CTC images (5% of all cells) including 555 for training and 139 for testing; 13472 non-CTC (95%) was down sampled to 555	contouring each CTC are needed for training; non-CTC are also needed to be segmented for training purpose	95.3% (sensitivity) 91.7% (specificity)	counting error: 5.4% (training) vs. 3.6% (testing)
S. Wang, et al. [16]	Apply YOLO-V4 on CTC counting	8 advanced-stage cancer patient samples; 6 normal samples	manual tight bounding box for each cell for training	an average precision (AP) of 98.63%, 99.04%, and 98.95% for cancer cell lines HT29, A549, and KYSE30.
M. Yu, et al. [24]	Segmentation results of all nucleis (CTCs and non-CTCs) from edge detection were used by a CNN for CTC detection	2300 cells from 600 patients with 1300 cells for training, 1000 cells fro testing (700 non-CTC cells and 300 CTC cells);	The Manual Interpretation of CTCs Counting is needed. CTCs and non-CTCs manual segmentation are needed.	The sensitivity and specificity of 90.3 and 91.3%.	No previous knowledge was adopted.
MG Krebs et al. [25]	Use the man-crafted features in classification.	699 cells in total, 80% for training. Not clear the number of CTCs vs. non-CTCs.	Manual labeling on the CTC types.	SVM with 10 features achieved the AUC of 0.99.
Satelli, et al. [26]	Crop CTC and non-CTC images are used for training a CNN to determinating CTCs	A total number of 120 cells (31 cultured cells and 89 WBCs) were tested.	Manually crop images of both CTC and non-CTC for training	No report on CTC counting errors.	It cannot handle two cells who are close to each other.
Shen, et al. [30]	Using computer vision preprocessing and ensemble generic deep learning with pretrained model from COCO data	mCTC with total 7623 images (5699 training and 1924 testing); CAF with total 702 images (455 training and 247 testing)	Two experts to do bounding box around CTCs manually, with the 2nd expert refining the labels from the 1st expert (high time cost of annotations)	Precision 94% recall 96% for mCTC	Precision 93% recall 84% for CAF
HI Method (This Paper)	Hybrid intelligence: using auto-segmented results and features with SVM for discriminating CTCs.	1070 CTCs for deep learning model training with another 323 CTCs for enumeration	Drawing a “weak” Box or rough circle around CTCs	Box Detection with Precision 0.98, Recall 0.92	SVM Enumeration with counting error 0.03.