Table 3.

Used algorithms and methods of data augmentation in tumor budding diagnosis of colorectal cancer.

Article	Data augmentation methods	Algorithms
Banaeeyan R. et al. Tumor Budding Detection in H&E-Stained Images Using Deep Semantic Learning	Horizontal reflection, vertical reflection, rotation, and scaling	- Encoder–decoder network with a pixel-level classification layer embedded in its last layer (an adaptation of the original SegNet model); - A total of 18 convolutional layers, each one followed by a batch normalization and rectified linear unit layer. The fully connected layer is not embedded in the architecture; - Every encoder has its own corresponding decoder and therefore the decoder block contains 18 layers. Binary TB class probabilities are obtained by feeding the output of the latest decoder layer into a softmax classifier.
Bergler M. et al. Automatic detection of tumor buds in pan-cytokeratin stained colorectal cancer sections by a hybrid image analysis approach	Median-filter to smoothen the image by reducing noise	- Classical methods like thresholding, filtering, and morphological operations for the detection of candidates. - AlexNet model for distinguishment between “true-positive” and “false-positive” tumor bud candidates.
Bokhorst J. M. et al. Automatic detection of tumor budding in colorectal carcinoma with deep learning	Random flipping, rotating, elastic deformation, blurring, brightness (random gamma), and contrast changes.	- VGGlike network with 2 configurations: one with 2 output classes (TB versus Background) and one with 3 output classes (TB, TG, Background); - L2 regularization and dropout layers added after the 2nd and 4th max-pool layer; - Multinomial logistic regression objective optimization (softmax), using stochastic gradient descent with Nesterov momentum.
Lu J. et al. Development and application of a detection platform for colorectal cancer tumor sprouting pathological characteristics based on artificial intelligence	–	- Faster RCNN developed by fusing the region proposal network (RPN) on the basis of Fast R-CNN; - VGG16 framework for feature extraction, RPN for feature region proposal and Fast RCNN network for boundary box classification and regression;
Fisher N. C. et al. Development of a semi-automated method for tumour budding assessment in colorectal cancer and comparison with manual methods	–	- The semi-automated method based on a binary threshold classifier built within QuPath (v0.2.3); - A pixel classifier created in QuPath to identify connective discrete areas of immunopositivity by combining image downsampling, stain separation using colour deconvolution, Gaussian smoothing, and global thresholding within a single step.
Pai R. K. et al. Development and initial validation of a deep learning algorithm to quantify histological features in colorectal carcinoma including tumour budding/poorly differentiated clusters	–	- CNNs were trained to segment each of 4 layers; - The first CNN segmented tissue into carcinoma (exclusive of TB/PDCs), TB/PDCs, stroma, mucin, necrosis, fat, and smooth muscle; - The second CNN segmented stroma into immature, mature, and inflammatory; - The third CNN segmented carcinoma into low-, high-grade, and signet ring cell carcinoma; - The fourth CNN identified TILs within the carcinoma layer.
Weis C. A. et al. Automatic evaluation of tumor budding in immunohistochemically stained colorectal carcinomas and correlation to clinical outcome	–	- 8-layer MatConvNet CNN model; - Color deconvolution to separate the background and foreground staining; - k-means clustering for thresholding.