Table 9.

General strengths and limitations of DL studies.

Author	Strengths	Limitations
Hou et al. [50] (2019)	Their approach generalizes significantly better to cancer cases without training data. Investigated the best quality without supervision cost.	Not generalized with mixed-quality image classification.
Shapcott et al. [101] (2019)	Reported cellularity, which has been linked to patient and prognostic indicators and other diagnostic.	Not handling small clusters or tumor polyps of up to five polyps in the stroma, which is associated with aggressive cancer.
Ben Hamida et al. [117] (2021)	Their method improved the results when treating with a sparsely annotated dataset	Techniques for accuracy not compatible with computational cost balance.
Liewa et al. [118] (2021)	Their approaches are robust enough to assist in CAD	Their classification system can misclassify images taken by colonoscopy/endoscopy according to the structure and image color characteristics, which are naturally irregular in the colon.
Pacal et al. [20] (2020)	Presented a comprehensive survey with all overviews.	Their model did not determine a common experimental setup and evaluation criteria.
Sikder et al. [119] (2021)	The method is pointedly precise, supported, and practical. Can detect malignant cells automatically. Their collection gives high accuracy, particularly after performing the algorithm of ML.	For large datasets, the used algorithm showed high complexity time. Low accuracy rate.
Kang et al. [104] (2019)	Used a strong object for detection CNN called Mask R-CNN. Utilized a successful ensemble model for combining the two masked approaches of R-CNNs with various backbone structures.	Less backbone structures. Less efficient segmentation, However, the successful ensemble method should be used with backbone structures.
Sornapudi et al. [105] (2019)	Successful detection and accurately used the segmentation method. The proposed approach showed better performance on the WCE video frames than images of colonoscopy.	Used the Etis-Larib dataset that does not produce efficient precision Training data is not sufficient for an accurate model.
Jia et al., [112] (2020)	Improved the residual learning and the feature pyramids. Developed the segmentation task of polyps.	Less stage integrations of PLPNet.
Zobel et al. [107] (2019)	Reduced the computation time. Detected too many FB areas.	The small training database for training a Mask R-CNN with a ResNet-101 backbone.
Ma et al. [108] (2019)	Overcome the problems of overfitting and gradient vanishing.	Not enough images for a training model.
Shaban et al. [64] (2020)	Well-suited for the CRC staging task.	Not efficient for digital images at the whole-patch level for the analysis of patient survival.
Blanes-Vidal et al. [109] (2019)	The used algorithm was able to determine the polyps’ similarity and determine the degree that is related to how true a match is. The used algorithm can be generalized.	High cost of information required on the location assessment and the polyp morphology. The detection algorithm was not efficient with the number of used images.
Wang et al. [120] (2020)	Achieved the effect of an automatic detection system for polyps, which was dependent on the DL for the detection rate for polyps and ADR.	The proposed system may be difficult to evaluate. Lack of external validity. False-positives rates were low. Fatigue level of participating endoscopies were not controlled for in this system, which considered this as an independent factor on ADR.
Mostafiz et al. [115] (2020)	Produce computer-aided system with great accuracy.	Small amount of FP and FN values.
Yuan et al. [111] (2019)	Model of DenseNet-UDCS was superior in accuracy of detection.	In the dataset, there are variances of small inter-class and unbalanced images and large intra-class differences.
Nadimi et al. [114] (2020)	The general rules are task-independent with less ambiguity for optimal feature selection. Better results compared with other state-of-the-art detection of polyps by a wide margin. Network predictions are given more interpretability.	Did not produce sufficient concrete interpretability.
Ozawa et al. [116] (2020)	Trained CNN presented a robust result for the detection and classification of CP.	This is retrospective research in a single association.