Ensemble learning [5] is a technique that relies on the combination of multiple machine learning algorithms to increase the performance of the model, and achieve a higher level of accuracy than would be achieved by using one of these algorithms separately. Algorithms mainly encounter two sources of error that prevent them from generalising beyond their learning set. These are Variance and Bias.

Variance is the error due to sensitivity to small fluctuations in the training set. A high variance can lead to overlearning. Bias is the error arising from incorrect assumptions in the learning algorithm. A high bias may be linked to a lack of relevant relationships between the input data and the predicted outputs (under-learning). To cover the maximum amount of information contained in the data and minimise processing errors, Ensemble learning methods are grouped into two main families: parallel methods and sequential methods.

Bagging or bootstrap aggregation (L. Kabari et al. [6] ) is a learning method where models are trained in parallel. The aim is to reduce variance. It involves sub-sampling the data, creating a data set for each model. To determine the final result, a vote is taken for classification, or the mean for regression.

To predict mortality associated with gastric cancer, but also the occurrence of any complication following surgery, C. Neto et al. [7] trained with the bagging method, a decision tree algorithm, simple logistic regression and a Random Forest. The study showed that the decision tree algorithm was the best technique for predicting mortality, with an accuracy of around 74%. Furthermore, the Random Forest algorithm showed the best results for predicting the possible occurrence of complications, with an accuracy of around 83%.

For K. Kaulanjan et al. [8] , in their study to predict biological recurrence after prostatectomy in prostate cancer patients, the Area Under the Curve (AUC) obtained were 0.65, 0.64 and 0.57 respectively for the Random Forest, Support Vector Machine (SVM) and k-Nearest Neighbour (KNN) models.

Decision Trees, Random Forest and support vectors machine algorithms were trained in parallel to identify long non-coding ribonucleic acid (RNA) as a diagnostic biomarker for gastric adenocarcinoma by Qun et al. [9] . The diagnostic capacity of these three models was evaluated by acquiring the area under the Receiver Operating Characteristic curve, sensitivity and specificity. The AUCs were 0.797, 0.981 and 0.983, and the specificity and sensitivity of the three models were 75.0% and 97.1%; 96.9% and 96%; 96.9% and 97.1%, respectively. The authors conclude that the machine learning models are able to diagnose adenocarcinoma of the stomach in the following decreasing order of accuracy: support vectors machine, Random Forest, Decision Trees.

The Random Forest algorithm is special in that it works on its own with the Bagging method. It is made up of several decision trees, trained independently on sub-samples of the training data-set. Each tree produces an estimate and a majority vote is used to obtain the final result.

C. Xu et al. [10] used a Random Forest to predict the survival status of patients with gastric cancer. External validation was performed and performance was assessed by the area under the ROC curve. The results showed an AUC of 76.6% using majority voting. The authors then decided to weight the voting process according to the accuracy tested for each tree instead of assigning the same weight as the original Random Forest (RF) model. The Random Forest algorithm with weighted voting gave an area under the ROC curve of 78.9%, an increase of 3% compared with the majority process.

Boosting is the Sequential ensemble Learning family. It helps reduce bias. The algorithms are run one after the other, and each one learns from the errors of the previous one. It improves the accuracy and predictive performance of models by converting several weak learners into a single strong learning model. Among the most common boosting techniques are Gradient boosting and Adaboosting.

Adaptive boosting (AdaBoost) [11] was one of the first boosting models developed. It adapts and attempts to self-correct with each iteration of the process. AdaBoost initially gives the same weight to each data set. It then automatically adjusts the weights of the data points after each iteration. It gives more weight to misclassified items in order to correct them for the next round. It repeats the process until the residual error, or the difference between the real and predicted values, falls below an acceptable threshold. Overall, AdaBoost is more appropriate for classification problems.

L. Fan et al. [12] used adaptive boosting (AdaBoost), linear discriminant analysis, and logistic regression classifiers to predict the presence of cancer cells inside a lymphatic vessel (lymph vascular invasion). Lymph vascular invasion is associated with metastases and poor survival. It is also difficult to diagnose. The data used was a combination of computed tomography images and clinical variables from patients with gastric cancer. A reference pathological examination was carried out to validate the prognosis. The AdaBoost technique obtained the best result with an AUC of 94%.

Gradient Boosting is similar to AdaBoost in that it is also a sequential technique. The difference is that Gradient Boosting does not give more weight to misclassified elements. Instead, it optimises the loss function by generating baseline learners sequentially, so that the current baseline learner is always more effective than the previous one. This method attempts to generate accurate results at the outset rather than correcting errors throughout the process. Gradient Boosting can help solve classification and regression problems.

W. Leung et al. [13] sequentially trained several machine learning algorithms to predict the risk of gastric cancer within 5 years of Helicobacter pylori eradication. The data consisted of clinical variables. Performance was measured by analysis of the area under curve (AUC) of the receiver operating characteristic. During a mean follow-up of 4.7 years, 0.21% of patients in whom Helicobacter pylori had been eradicated developed gastric cancer again. The gradient boosting technique performed well in terms of prediction, with an AUC of 97%. This result was superior to that of conventional logistic regression (AUC of 90%), applied independently.

The applications of ensemble learning methods in oncology, developed in this section, are recapitulated in Table 1.

Artificial neural networks (ANN) are algorithmic imitations of the functions of the human brain. Like biological neurons, they receive input information, process it through a series of operations and produce output. Once properly trained, artificial neural networks can learn from themselves and constantly update themselves to provide increasingly accurate results.

The perceptron [14] is historically the first artificial neural network model. It has two layers of neurons: an input layer and an output layer. The two layers are directly connected. This type of algorithm is useful for the linear classification of a set of information into two distinct classes (present/absent, alive/dead, benign/malignant, etc.). But given the availability of massive amounts of data, its unstructured nature and the complexity of the relationships between characteristics, neural networks are being deployed in greater depth.

Previously, image analysis has been based principally on features such as colour, brightness, shape, texture pattern and other distinguishing characteristics. However, this type of analysis is limited by image rotation, lack of brightness, adjacent or angled views of the object, or image blur [15] .

Deep Learning involves introducing one or more hidden layers between the input and output layers in order to perform intermediate operations and unearth buried relationships and relevant features. Depending on the structure of the data, specific neural networks are developed.

For example, Convolutional Neural Network [16] is particularly well adapted to image processing and is widely used in cancer research. It uses convolution operations to extract features from images by applying filters that detect characteristics in the image. Convolution layers are usually followed by pooling layers that reduce the size of images while retaining key features. With their self-learning capability, the programme works without any indication of what is required.

Neural networks have made a significant contribution to cancer diagnosis, notably for the detection of genetic types, subtypes and mutations [17] , the identification of rearrangement status in prostate cancer adenocarcinoma [18] , and the evaluation of pelvic echography images [19] . Some studies have reported that the performance of neural networks surpasses that of medical specialists in the fight against cancer.

Lianlian Wu et al. [20] have developed a system using a convolutional neural network to detect early gastric cancer. Images of different parts of the stomach were collected to train the network to monitor blind spots and automatically cover suspected cancerous areas. The results were compared with the findings of an endoscopy team. For locating cancer cells, the network achieved 90% accuracy, on a par with endoscopy experts. The model also achieved 92.5% accuracy in classifying tumours as malignant or benign, surpassing the performance of the experts.

L. Xiangchun et al. [21] are using a convolutional neural network to diagnose thyroid cancer. The data consisted of echography pictures of thyroid cancer patients and healthy controls. A clinical diagnosis of the training set was performed by 16 qualified radiologists and a reference pathological examination was carried out for confirmation. The model achieved almost the same sensitivity as the qualified radiologists (93.4% versus 96.9%), but much improved specificity (86.1% versus 59.4%), in identifying patients with thyroid cancer.

The authors subsequently included serological markers in addition to echographic images in a new study [22] with the same objective. The previous model was refined and now consists of two parallel branches. One branch takes the input image and outputs a vector F representing the input image. The other branch takes the abundance of serological markers as input and outputs a G vector, a learned characteristic of the serological markers. The item-by-item summation of F and G was considered the integrated multimodal feature. Labels obtained from pathology reports were used as the diagnostic reference. The performance of the new model was improved compared with that of qualified radiologists (AUC: 95% versus 89%).

Neural networks have also been applied to cancer prognosis.

Nakahira H et al. [23] developed a convolutional neural network using endoscopic images. The aim was to stratify the risk of gastric cancer. The model was trained using stochastic gradient descent. The system identified three distinct groups of patients: low, moderate and high risk. Endoscopic examination and serum antibody tests were performed to validate the results and the prevalence were: 2.2%, 8.8% and 16.4% respectively for each group, with significant differences (p = 0.0017).

L.Q. Zhou et al. [24] trained a convolutional neural network to predict the occurrence of clinically negative lymph node metastases in patients with primary breast cancer, based on echographic images. A reference pathological examination was performed. The model achieved a sensitivity of 82% compared with 63% for expert radiologists.

The applications of neural networks in oncology, developed in this section, are resumed in Table 2.

Transfer learning [25] is a machine learning technique that transfers to a model the knowledge acquired by another model previously trained for a similar task with sufficient data. The pre-trained model is called the “source” and the model receiving the transfer is the “target”. The weights and biases learned by the source are then transferred to the target. Transfer learning can be useful when the data available for the target task is limited, or to reduce the time and resources required for training. This method is increasingly used in the fight against cancer, where it can improve the accuracy of tumour classification according to type, stage or aggressiveness [26] [27] , tumour detection [28] , and cancer prognosis [29] .

DenseNet networks [30] have an architecture called densely connected convolutional network. They are renowned for their prowess in extracting optimal features. In their study to extract relevant features for the identification of primary gastric cancer or lymphoma, B. Feng et al. [31] used DenseNet121 parameters on Whole slide images. The relevance of the factors selected was assessed in comparison with those selected by a team of expert clinicians. A logistic regression with the selected factors was used for validation and obtained an AUC of 96%, higher than the results of the clinical model.

A. Saber et al. [32] have developed a model based on transfer learning to provide effective support for the automatic diagnosis of breast cancer suspects.

Image features are extracted using a pre-trained architecture: VGG16 [33] . VGG16 is a convolutional neural network trained on a collection of more than 14 million images belonging to 22,000 categories. Experimental results show that the features selected enable breast cancer to be diagnosed with 98.96% accuracy.

ResNet-50 [34] is a convolutional neural network with 50 layers of depth. This model introduces residual connections. Unlike standard convolutional neural networks, which have a linear architecture (a stack of layers in which each output is connected only to the next layer), in a residual network the output of previous layers is connected to the output of new layers, so that they are both transmitted to the next layer. This architecture enables the creation of very deep neural networks, with greater precision, because they are able to extract more information and thus perform a more advanced analysis of pictures.

Z. Yafang et al. [35] used the performance of ResNet50 to extract features from echographic images to establish a prognostic model for hepatocellular carcinoma. The importance of the selected factors was compared with that of clinical parameters using a survival analysis. The results showed specificity (81.0% versus 38.1%) and accuracy (78.8% versus 51.5%) compared with the clinical model. In addition, microvascular invasion was identified as having a very poor prognosis for overall survival, with a Hazard Ratio of 6, and for relapse-free survival, with a Hazard Ratio of 3.3.

The applications of transfer learning in oncology, developed in this section, are summarized in Table 3.

Unsupervised approaches play a major role in making accessible and explainable the information contained in masses of unlabelled data, with no notion of output values. They make it possible to learn the characteristic structures and relationships buried in the data. These algorithms are left to their own devices to discover and present the interesting structure of the data. Unsupervised techniques are used to perform clustering based on similarities or differences in the data, in order to focus the search on significantly representative sub sets. Analysis of genomic data is also made possible by reducing dimensionality.

P. Apostolou et al. [36] conducted an unsupervised clustering analysis based on hierarchical and k-means algorithms to discriminate between colon cancer and normal conditions as well as between different types of cancer. The data

consisted of blood sets from colon cancer patients, normal donors and commercial cancer cell lines representing different types of gastrointestinal cancers. Gene expression analysis was performed for more than fifty genes. The Euclidean distance metric was used to highlight dissimilarity between the data. The distribution of the data was evaluated using the Kolmogorov-Smirnov test. The study effectively separated: 1) colon cancer cell lines 2) normal sets 3) stomach, pancreatic and liver cancers.

J. Gal et al. [37] , evaluated several unsupervised approaches to clustering according to survival in breast cancer patients, exploiting metabolomic signatures. These were Principal Component Analysis (PCA), k-means, Sparse k-means, Single-cell Interpretation via Multi-kernel Learning (SIMLR), k-sparse and Spectral clustering. The results show that three optimal groups (favourable, intermediate and unfavourable) were selected. In terms of processing times, PCA and k-means were the fastest and K-sparse was the longest. SIMLR and k-sparse methods were the most discriminants with an average silhouette value of 0.85 and 0.91, respectively. Survival analysis revealed a significant difference in predicted 5-year overall survival between the 3 groups. This approach shows the possibility of stratifying breast cancer patients using metabolomic signatures with unsupervised approaches.

Dimension reduction allows the study to focus on a finite number of significant and representative factors. For example, it is very difficult to analyse genomic data with tens of thousands of dimensions. Principal component analysis reduces dimensionality by converting a set of correlated variables into a set of principal components (i.e., linearly uncorrelated variables) using an orthogonal transformation.

Gene expression data from several cancers, including gastric cancer, were exploited in the study by J. Xie et al. [38] . The aim was to identify optimal biomarkers for the detection of cancer cells. Principal component analysis was applied. All the characteristics are displayed in a two-dimensional space with the standard deviation defining discernability on the abscissa and the cosine similarity defining independence on the ordinate. By calculating the area bounded by the coordinate lines and the principal axes, the optimal features are located in the top right-hand corner of the space. To validate the results, SVM models are built with the detected biomarkers. The results show 100% accuracy for classification into malignant or benign tumours. As a result, the proposed technique is powerful for detecting subsets whose features are relatively independent of each other.

F. Yang et al. [39] studied 3385 differentially expressed genes obtained from single-cell RNA sequencing data of gastric cancer specimens. The aim was to find biomarkers with high sensitivity and specificity to accurately assess patient prognosis. Principal component analysis was applied to the dimension reduction process. Three cell subsets were identified: gastric cells, plasmacytoid dendritic cells and T-memory cells. In addition, the fatty acid binding protein (FABP1) was identified as the pivotal gene in gastric cancer progression. FABP1 was found to be closely related to long-term survival and age at diagnosis of patients.

The applications of unsupervised approaches to oncology developed in this section are summarized in Table 4.

Model mixture consists of combining different machine learning techniques, in order to exploit the advantages of each to create a more complete and more effective model. Faced with ever-growing volumes of data, the complexity of the relationships between features, and the performance requirements imposed on researchers, a single learning model may be limited in its ability to cover all the information contained in the data. The mixture strategy is like a distribution of tasks. The algorithms used each process a specific task according to its efficiency and the structure of the data. When applied to cancer data, the mixtures of model’s strategy have produced very satisfactory accuracies.

O. Iizuka et al. [40] developed a neural network model for the classification of histopathology pictures of the stomach and colon. In the feature extraction step, they trained convolutional neural networks for each organ separately. These networks were based on the pre-trained inception-v3 architecture [41] , removing the fully connected classification layer. Each image contained an arbitrary number of extracted features. Knowing that a Recurrent Neural network (RNR) [42] can take a sequence of arbitrary length and produce a single output, all the extracted features are used as input to an RNR with two Long Short-Term Memory (LSTM) layers. The final model achieved good classification performance with an AUC of 96%, compared with 86% for expert pathologists and only 41% for students.

H. Zhang et al. [43] combined several models to detect gastric cancer from histopathological images. First, pre-trained neural network architectures were

used for image segmentation. Next, feature extraction is facilitated by clustering using the k-means method. Finally, different classification algorithms were trained with the selected factors. The results of this study show that using the convolutional network “U-Net” for image segmentation, feature extraction based on graphs and finally the Support Vector Machine classifier, gastric cancer can be detected with an accuracy of 94.29%.

A. Dongyao et al. [44] and A. Moloud et al. [45] have also combined convolutional neural network with Support Vector Machine, and obtained accuracies of 99.3% and 100%, respectively, to provide accurate diagnosis of cervical cancer and breast cancer, respectively.

In order to predict the occurrence of lymph node metastases, Z. Xue et al. [46] conducted a retrospective study using clinical and pathological data from patients with stage 1 and 2 gastric cancer who had undergone surgical treatment. Univariate logistic regression was used to detect variables significantly affecting metastasis. These variables were given as input to a Multi-Layer Perceptron (MLP) neural network to establish a prognostic model. The results were compared with the real postoperative state. Logistic regression showed that the platelet/lymphocyte ratio, systemic immune inflammation index, tumour size and clinical stage were closely related to lymph node metastasis. On the basis of these variables, the MLP model achieved an AUC of 75%, compared with 54% for post-surgical anatomopathological staging and only 30% for pre-therapeutic clinical staging. This technique makes it possible to predict metastases in the early stages of cancer.

K. Priyanka et al. [47] developed a hybrid model to predict the complete response of breast cancer before the start of chemotherapy. A convolutional neural network is used for image feature extraction. Mann-Whitney tests were used to evaluate the features and their relevance. Then, the selected factors were provided independently as input to different machine learning classifiers including (Random Forest, SVM, DT, KNN etc.). Classification performance was evaluated using accuracy and the area under the receiver operating characteristic curve (AUC). The results show that the combination of the CNN and the KNN classifier performed best with an accuracy of 99.8% and an AUC of 100%.

The applications of model mixing in oncology developed in this section are summarized in Table 5.

Recent developments in applications based on artificial intelligence offer a multitude of possibilities for exploiting data and improving levels of precision. Strategies based on ensemble learning, deep learning, transfer learning and unsupervised approaches make algorithms increasingly flexible in adapting to data and finding solutions to research questions.

Choosing the right model for a classification or regression, in the presence of labelled data, is made easier by the bagging technique, which enables several models to be trained in parallel. Boosting also means that algorithms can be run sequentially to produce a strong learning model.

Deep learning enhances the self-learning capacity of neural networks. Intermediate calculation layers enable complex and unstructured data volumes to be exploited in depth. Convolutional Neural Networks are particularly well suited to image processing, and are widely used in cancer research.

Transfer Learning can be used to overcome the lack of data in some contexts. The knowledge of a pre-trained model with sufficient data can be used to solve a similar problem with few data of the same dimensions. Algorithms such as ResNet, DenseNet, VGG16, etc. are reused in new studies.

Unsupervised approaches are used to learn structures and characteristic relationships buried in the presence of unlabelled data. These algorithms are left to their own mechanisms to perform clustering based on data similarity or dimensionality reduction.

Hybrid models involve using several algorithms to solve a problem. Perfect coordination is established and each model executes a specific task according to its efficiency and the structure of the data.

All these techniques help to considerably reduce the time and resources needed to train the models, and enable high levels of accuracy to be achieved.

Despite the reported success of AI, there are still a few limits to the widespread use of artificial intelligence solutions.

Data bias is a problem faced by artificial intelligence algorithms. To illustrate, let’s take the case of the Cancer Genome Atlas (TCGA), which is considered to be the largest repository of diverse cancer datasets. However, it is mainly composed of data from white individuals of European descent [48] . Consequently, the results proposed by the algorithms may be subject to errors due to selection bias [49] .

Deep learning has the reputation of being a “black box”. Neural networks do not provide explanations for their results, which limits the analysis of existing phenomena [50] . Non-interpretability affects understanding and confidence in proposed solutions. Fortunately, some researchers are tackling this problem, with promising results [51] [52] .

Access to sufficient high-quality training data is also a major obstacle to the adoption and widespread use of artificial intelligence solutions [53] . Although transfer learning seems to offer the beginnings of a solution to this data shortage, it is currently limited to image exploration, for which data of the same structure and dimensions are readily available. For numerical variables, on the other hand, the problem of standardization is an obstacle. In fact, there is no standard codification or unit of measurement.

Synthetic data generation in turn offers a judicious approach to this situation. It consists in generating artificial data from existing real data, based on their distributions and correlations between attributes [54] .

This synthetic data can then be used for a variety of purposes, such as training machine learning models, validating data analysis methods, or preserving the confidentiality of sensitive data.

In order to allow Random Forest algorithms to continue incrementing without resorting to old data at each iteration, J. Gonzalez et al. [55] proposed a method consisting in the generation of synthetic data using the normal distribution.

H. Akrami et al. [56] have also developed a variational autoencoder for tabular datasets with categorical and continuous features that is robust to outliers in the training data.

Lei Xu et al. [57] have designed a CT-GAN (Conditional Tabular-Generative Adversarial Network) that relies on mode-specific normalization for continuous columns, architectural changes, and consideration of data imbalance for discrete columns.