The absence of symptoms during the early stages of leukemia makes it challenging for practitioners to predict the occurrence of cancer [117]. Typically, leukemia detection relies on the use of blood cell image classification methods [29]. The introduction of AI-based models is intended to enhance identification accuracy and provide early prediction of potential spread within the human body to help increase the chances of patient recovery and survival.

No standardized leukemia screening tests have been proven reliable and efficient enough to identify the presence of blood cancer in its early stages [83]. The causes of leukemia remain unknown and a diagnosis cannot be confirmed until the initial symptoms develop. Thus, there is ongoing research on imaging techniques, screening methods, and ways to increase the contribution of AI to provide a better understanding of the root causes of this condition and thus enable expert hematologists to identify and diagnose blood malignancies accurately [56].

For many diseases, diagnosis is the most crucial stage in the illness pathway [85]. As the field of hematologic research is expanding in an unprecedented manner, the emergence of AI provides an opportunity to overcome the challenge of handling large amounts of imaging data, improve the efficiency and quality of hematologic pattern identification, and provide a clear understanding of suitable therapies following a precise diagnosis [87].

AI integration into blood cancer treatment has enabled many advanced practices to deliver timely and efficient therapy [20]. Several techniques were used to identify the doses and combinations of drugs required for each patient as functions of their health status, age, and other essential factors. Table 3 classifies the 131 journal papers based on their corresponding pathway stage.

While there has been more emphasis in the literature on the diagnosis stage, which represents 39.7% (52/131) of the total journal papers collected, treatment and prediction are the least investigated pathway stages with only 13.7% (18/131) and 20.6% (27/131), respectively. This is mainly due to the difficulty in identifying crucial biomarkers that can be used as discriminative feature variables to predict the disease before the onset of symptoms. Similarly, in the case of treatment, as detailed in later sections of this paper, most of the AI-based models [9,42,75] reported in the treatment phase utilize gene expression profiles for model development, which are relatively complex. Hence, more research is needed to investigate the feasibility of easily available biological features for the prediction of treatment-related disease relapse and patient survival.

Childhood ALL is a malignant cancer that is the leading cause of pediatric cancer mortality, and around 20% of the children fully treated end up having a recurrence [131]. Hence, it is crucial to predict relapse to deal with the multiple risk groups accordingly. For better management and follow-up planning, Pan et al [131] introduced an ALL relapse prediction model based on ML algorithms that help classify patients with ALL into their appropriate risk categories. In the model selection process, 103 clinical variables were used to train 4 classification algorithms, random forest (RF), decision tree, support vector machine (SVM), and linear regression, to distinguish relapses from nonrelapses in the 3 clinically predefined risk categories: standard-, intermediate-, and high-risk levels. While Pan et al [131] built a model to predict disease relapse, Hauser et al [144] studied the possibility of predicting CML prior to diagnosis using only CBC test results and ML algorithms such as XGBoost and LASSO algorithms on 1623 patients with a definitive CML status. The variables used in the study included laboratory CBC test results, patient demographic features such as their age and gender, and patient encounter information (the number of patient visits to outpatient clinics, etc.). A forward feature selection process was employed to measure the predictive performance of the most potential predictors. The data set was then divided into 7 subsets, in which time of diagnosis was set as a patient baseline, and the 6 remaining sets corresponded to the different time periods preceding the diagnosis test. Interestingly, variable selection yielded different features for inclusion in the models depending on the data collection interval.

The performances of the chosen classifiers in [131] were evaluated by a 10-fold cross validation in each of the 100 training sets. However, the suggested approach is considered insufficient for internal validation that requires at least 50 repeats [145]. By contrast, the chosen data set in [144] was divided into 2 distinct groups: train/validation and test groups. While the latter split-sample validation approach seems reasonable and justifiable to use in this case given the large sample size, potential drawbacks may arise, and several aspects still need attention throughout the application. For instance, as the sample split was performed fully at random, substantial patient imbalances might have occurred with respect to the distributions of predictors and the output. Moreover, 20% was used for model assessment leading to a potential biased evaluation of the model’s performance [145].

Furthermore, it is well known that patients with leukemia often suffer from health issues due to frequent infections, which can lead to death if it is not detected early. Toward this end, Agius et al [55] investigated the risk of infection due to a weakness in the immune system or a cytotoxic treatment immediately after CLL diagnosis by developing the CLL Treatment-Infection Model (CLL-TIM). For each patient, the prediction point was set at 3 months after their diagnosis, and the target output was the 2-year infection risk or CLL treatment. After excluding 74 patients who died and 373 who initiated their treatment before the prediction point, the study cohort’s final size corresponded to 3729 patients. In contrast to Hauser et al [144], Agius et al [55] employed stratified sampling to maintain class distributions and compensate for the 52% International Prognostic Index for CLL (CLL-IPI) missing variables, by dividing the data set into 65% training set and 17.5% for each of the test and internal validation sets. Using 7288 features resulting from a collection of variables from different sources, comprising baseline variables at the time of diagnosis including age, gender, etc.; routine laboratory tests; microbiology findings; pathology reports; and diagnosis codes for all patients, the CLL-TIM ensemble algorithm was composed of 28 ML models that could identify patients at a high risk of infection to increase their chances of survival. The model was then validated on both internal and independent external test cohorts, and exhibited interesting performances, surpassing the CLL-IPI.

Hassouneh et al [50] suggested the use of deep neural networks (DNNs) to predict survivability of patients with leukemia to boost the psychological state of patients and enable physicians to arrange the proper treatments for different cases. The final DNN structure encompassed 6 hidden layers, with 45 hidden neurons in each corresponding hidden layer, and a dropout activation with 25%. While Hassouneh et al [50] used patient records and attributes of patients with leukemia for modeling, Boldú et al [134] relied on a set of 731 blood smear images to predict initial patient diagnosis. For data set limitation purposes, 4 convolutional neural network (CNN) models were pretrained on a larger data set and used for the aforementioned task. Their respective performances were compared and the architecture achieving the best outcome was trained further. Next, these pretrained CNNs were fine-tuned to match the type of data that are fed into the model. Finally, the proposed ALNet model was able to distinguish, on a first level, between healthy and abnormal blood cell images. On a second level, it was able to identify whether the blasts were myeloid or lymphoid. After a thorough evaluation process using 5-fold cross validation and the hold-out (80%/20%) approach with 470 iterations, ALNet demonstrated interesting results and was chosen as the best architecture for modeling. One strength of this study is that the 5-fold cross validations were approximately balanced, and despite the random split, the data from the same patient smear were maintained within the same fold. Alternatively, the proposed DNN algorithm in [50] was trained and evaluated using 2 methods: 10-fold cross validation and ensemble method. Although the 10-fold cross-validation method is known to generate more stable results, the number of repetitions is quite crucial, which was absent in this study [145].

Overall, it is very clear that all the aforementioned algorithms succeeded in predicting the different aspects and repercussions of the disease. Nevertheless, no existing literature has yet examined the root cause of the malignancy by predicting the possibility of patient infection. Table 4 summarizes some studies in the prediction phase alongside their objectives, the data sets used in the studies, the methodologies followed, the performance of the models applied, their strengths and weaknesses, and the validation approach used, where available.

Initial leukemia screening and its efficient diagnosis require a deep and thorough image analysis process. As opposed to traditional manual screening, automated leukemia screening is a novel approach that minimizes human interaction and provides more accurate clinical information by using blood smear images to identify ALL automatically [17]. The automated screening process is considered challenging due to the leukocyte localization and region extraction phases, which are generally obtained via background removal and separation of surrounding blood components that might distort the overall detection process. For this reason, many studies have employed techniques such as principal component analysis as a filter to identify any features that do not bring any important information to the classification process [115] to enhance detection accuracy. Similarly, Chebouba et al [32] used a meta-heuristic stochastic local search technique to select the most important genes and proteins to be used in the RF-based classification of patients with AML.

CML consists of 3 sequential phases that change based on the patient’s status and can progress to more severe phases if timely treatment is not provided. This makes CML phase identification very crucial, as different phases require separate treatments and medical regimens. In the chronic phase, less than 10% of the cells in both the blood and bone marrow are blasted. The severity and persistence of the aforementioned phase depend mainly on the consistency of the therapy followed. If the chronic stage is neglected and the patient does not receive timely and effective treatment, the condition can deteriorate to reach an accelerated phase where the blast count increases to around 10%-19%. Similarly, if the patient’s condition declines with no appropriate medical intervention, the percentage of blasted WBCs doubles to reach around 20% or more. At this stage, the patient’s state is considered uncontrollable, and the patient starts to exhibit symptoms such as fever, weakness, and weight loss [33]. As CNNs were proven to be efficient tools for accurate image recognition, Khosla and Ramesh [33] suggested using the latter to classify different CML images into their respective phases. Similarly, Togacar et al [71] used CNNs to separate WBC images into their 4 subclasses: eosinophil, lymphocyte, monocyte, and neutrophil. While the aforementioned CNNs are successful in identifying the most important features in images with no human supervision, they generally require large training data to achieve high performance, and their employment is regarded expensive in terms of both time and training. For this reason, while many studies tend to employ image augmentation to the existing data set to create larger samples by slightly changing the existing collected images [33], others implement the concept of transfer learning to overcome the training data shortage. The idea behind the latter is to leverage the power and knowledge of a pretrained model to apply it on a new similar task. For instance, Sahlol et al [3] proposed a novel approach consisting of a hybrid model and combined CNN feature extraction using the solid architecture of VGGNet that was pretrained on ImageNet, to separate malignant cells from benign ones. Similarly, Li et al [127] developed the globally optimized transfer deep-learning platform with multiple pretrained CNNs (GOTDP-MP-CNNs). This DL platform is composed of 17 CNNs able to classify pathologic images into human diffuse large B-cell lymphoma and non–diffuse large B-cell lymphoma. Table 5 summarizes some of the collected studies in the screening phase.

To address the challenge of manually detecting blasted cells, Dasariraju et al [54], Inbarani et al [66], Abedy et al [29], Jagadev and Virani [34], and Dharani and Hariprasath [31] used medical images of healthy and malignant samples to automatically identify the leukemic types and subtypes. While Dasariraju et al [54] applied an RF algorithm as an approach to differentiate between abnormal and healthy leukocytes, and classify immature leukocytes into their 4 subtypes, Inbarani et al [66] discussed the implementation of a novel sophisticated approach to identify ALL blast cells via the histogram-based soft covering rough K-means clustering (HSCRKM) segmentation algorithm. The latter is a hybrid-clustering technique that combines the strengths of both the soft covering rough set and the rough K-means clustering. Nevertheless, one main limitation of the HSCRKM segmentation technique is that it is not suitable for multiple color images because the latter increase the processing time due to an increase in the peak values of the histogram. To enhance image representation and prepare a clean input for the classification model, Jagadev and Virani [34] applied SVM on 220 blood smear images of healthy individuals and patients with leukemia to identify the 4 leukemic subtypes (AML, CML, CLL, and ALL) using both K-means clustering and hue, saturation, value color–based segmentation techniques. Similarly, Dasariraju et al [54] performed a set of morphological imaging modifications and preprocessing techniques to segment the nucleus and cytoplasm and overcome the difficulty of blood image detection. To reduce data dimensionality, the model’s speed, and processing time, the most relevant features were obtained using feature extraction techniques and the resulting output was passed on to the classifier [34,54,66]. Alternatively, Abedy et al [29] chose to employ the histogram of oriented gradients for feature extraction, the Gaussian filter for noise elimination, and the Sobel kernel for image filtering. Furthermore, Dey and Islam [49] adopted the principal component analysis technique followed by grid search for hyperparameter tuning, which significantly decreased the number of components from 7129 features to only 6 important parameters. This data transformation technique did not only reduce the computational time and made it faster, but it also helped giving better results. Alongside the usage of imaging techniques, many studies employed other techniques; for example, Moraes et al [133] suggested the usage of flow cytometry data for distinguishing leukemia/lymphoma, and Mahmood et al [143] directed their research to focus more on identifying the most discriminatory features for CLL using patient laboratory test results, demographic parameters, and training a Classification and Regression Trees model on 94 pediatric patients, which was evaluated using 10-fold cross validation. Moreover, both Dharani and Hariprasath [31] and Jagadev and Virani [34] used SVM to classify leukemia and its subtypes, while Paswan and Rathore [28] used K-nearest neighbors to separate blasted blood cells from normal ones and classify them further into either AML or ALL using a value of K=4. By contrast, Moraes et al [133] suggested the implementation of decision tree as an ML-based technique for distinguishing leukemia/lymphoma, where a binary classification between healthy and immature leukocytes was performed with an 80%/20% data split, followed by a subclassification of immature leukocytes into their respective 4 types using a 70%/30% split, and several combinations of hyperparameters were evaluated during a 5-fold cross validation. Conversely, Abedy et al [29] chose to employ logistic regression as a classifier to identify the shape of the leukemic cell from microscopic blood images, while Dey and Islam [49] conducted a study to detect patients’ leukemia type based on their gene expression information using the RF algorithm and 2 other algorithms, XGBoost and artificial neural networks (ANNs), and Dasariraju et al [54] used RF to perform a subclassification of immature leukocytes into their respective 4 types.

With the objective of optimizing leukemia diagnosis, some studies made use of genetic features. For example, Rodrigues and Deusdado [63] suggested the application of a kernel logistic regression that classifies gene expression data using meta-learners to select the most relevant attributes before classification. The data set used for training comprised the 2 types of leukemia (ALL and AML) and a set of gene features. Pearson correlation and chi-square statistic were the 2 approaches used on meta-learners to assess attributes. Then, all models used 10-fold cross validation resulting in an identification of 12 common genes. Other studies, such as that by Al-Dulaimi et al [23], focused on current practices, techniques, and challenges in digital hematology detection of WBCs and their components (nuclei and cytoplasm) using hematologic microscopy images. In addition to analyzing the growing trends in computer-aided diagnosis applications, the review highlighted the main challenges associated with the use of CNNs in terms of both high computational time and costs to classify images and detect abnormalities. This gave rise to the use of transfer learning and enhancing optimization techniques, such as the bare bones particle swarm optimization algorithm used by Srisukkham et al [12] to extract the most informative features and enhance the classification accuracy of the lymphocytic cells into either normal or blasted. Likewise, Miyoshi et al [122] directed their work toward enhancing lymphoma diagnosis by classifying histopathological lymphoma images using a DL model. The aim of the study was to evaluate the performance of the suggested automated DL model and compare it with that of a traditional manual hematopathology detection procedure using CNNs. In this study, each test set comprised a total of 100 image patches, and the rest was randomly divided into 5 separate groups. During each repetition for 5 iterations, 1 group was kept for validation to evaluate the classifier performance for every epoch, while the other 4 groups were used for training. The number of epochs used in this study was 30. Similarly, Shafique and Tehsin [125] suggested the deployment of deep CNNs to detect the ALL type and its corresponding subtypes, while Zhao et al [60] proposed turning the captured raw multiparameter flow cytometry data into a 2D image by means of a self-organizing map (SOM) to analyze and classify them using the aforementioned algorithms. An SOM is a map of neurons that relies on unsupervised learning, in the sense that human intervention is not necessarily required. This model is used in many applications and has strong generalization abilities. The model was trained for 15 epochs, at a learning rate of 0.001 using an Adam optimizer function. The evaluation of classification accuracies used a 10% validation split, which was also employed for network architecture optimization. Then, the model performance was evaluated using the hold-out test set. In the same context, Sipes and Li [4] attempted fine-grained image classification for ALL diagnosis and compared the accuracies of CNNs and other models that used specific hand-selected features, while Vincent et al [37] proposed a leukemia classification that could be performed on 2 levels. The first process was applied to well-segmented nuclei extracted from 100 blood smear images. In this step, 90 samples were used for training and 10 were kept for validation. The cells were classified into normal and abnormal. The second step consisted of feeding 5 features extracted from abnormal images into a second classifier that split the images into ALL and AML types accordingly. Table 6 summarizes some of the studies in the diagnosis phase based on their objectives, data sets used, methodologies, and performance characteristics.

Complete remission (CR) refers to the disappearance of all signs and symptoms of an illness [51]. However, a significant proportion of patients report disease relapse after therapy and complete disease recovery. In this regard, Gal et al [75] proposed an ML technique that uses gene expressions to predict the likelihood of CR in patients with AML who previously received therapy. The 473 collected samples were divided into training and testing and were fed into the 3 classifiers for feature selection using a 5-fold cross validation. To select the most significant genes that clearly mark the difference between the state of CR and non-CR, a statistical t test was performed in each fold for each method. For further gene feature selection and performance enhancement, the results were compared with those of 3 algorithms: randomized LASSO, recursive feature elimination, and hill climbing. It was proven that cancers that appear pathologically identical do not necessarily exhibit the exact response to similar drugs. To overcome the challenge of high demands for personalized or patient-specific medicine, Lee et al [9] proposed using a gene-expression profile and in vitro drug sensitivity data to spot molecular markers that explain this patient-specific drug response. The data set used comprised 160 chemotherapy drugs and inhibitors for 30 patients with AML. The gene-drug association was identified using the MERGE algorithm, which utilized gene characteristics such as a novel mutation, expression hubness, known regulator, genomic copy number variation, and methylation. The model testing was done in 2 different ways: the first approach used 2 batches containing 12 patient samples for training and 12 different samples for validation, respectively. The second approach used a leave-one-out cross validation to test the predicted drug sensitivity for 30 patient samples. The latter cross-validation method is regarded as reasonable because it prevents the high computational training cost and time owing to the small sample size and is much less biased than using a single test set because the process fits the data set consisting of n-1 observations repeatedly [145]. Although having a limited data size permitted the obtention of efficient validation results, identifying gene-drug associations is deemed to be a challenging process in that case. However, the proposed algorithm was successful to prioritize genes based on the multidimensional data on their potential to drive cancer. Consequently, upon comparison with other alternate methods (ElasticNet, multitask learning, Pearson P value, and Spearman P value), MERGE exhibited the best gene-drug association result.

In their review and analysis of current AI applications for hematologic disorders’ treatment, Muhsen et al [24] presented a study that was performed with a set of patients who underwent allogeneic hematopoietic cell transplantation to predict the development of acute graft-versus-host disease using ANNs. After comparing the performance acquired by ANNs and the results achieved by logistic regression, ANNs were found to predict the presence of graft-versus-host disease significantly better. However, among the limitations remaining to help reach optimal ML results is the limited data input from patients that could be further enlarged to include biologic and genetic factors, for instance. Moreover, the employment of several sampling techniques such as random oversampling, synthetic oversampling, and remote under sampling could help improve the ML models’ accuracies in predicting treatment-related mortality in allogeneic hematopoietic cell transplantation.

Similarly, Lyu et al [42] used ANNs to classify the progress and change in gene expressions and peripheral blood mononuclear cells before treatment and after starting therapy. Table 7 shows some extracted literature focusing on the hematology treatment phase.