Table 6.

Study analysis for journal publications on the diagnosis phase.

Reference	Objective, data set, and methodology	Performance and remarks
[60]	Objective: Classification of mature B-cell neoplasm Data set: 20,622 routine diagnostic samples from Munich Leukemia Laboratory Methodology: CNN-SOMa transformation	Performance: Accuracy: 95% Strengths: Large data set High accuracy Limitations: Nonuniform distribution of misclassifications due to similarity in flow cytometric profiles Validation: 10% validation split
[54]	Objective: Detection of immature leukocytes and their classification into 4 types Data set: Images extracted from a publicly available data set at The Cancer Imaging Archive Methodology: Random forest algorithm	Performance: Accuracy: 92.99% Strengths: High precision results for each class Limitations: High number of false positives leading to low precision and specificity Validation: 5-fold cross validation
[49]	Objective: Identification of the leukemia type based on patient genetic expression Data set: A sample of 7129 genes that represent the genetic expressions of 72 people from Kaggle Methodology: XGBoost, artificial neural networks, and random forest algorithm	Performance: Random forest accuracy: 80.8% XGBoost accuracy: 92.3% Strengths: Use of principal component analysis for dimensionality reduction and faster computation Use of grid search for the best hyperparameter selection Limitations: Small data set (72 people) Validation: Internal validation (65%/35% split)
[135]	Objective: Classification of lymphocytic cells Data set: The ALL-IDB2 Database Methodology: bare bones particle swarm optimization–based feature optimization	Performance: Accuracy: 94.94%-96.25% Strengths: A good performance on capturing prognostic chronic myeloid leukemia markers by the model Limitations: Challenge of capturing relationships between data types with no information loss in clinical clustering Validation: Validation on an external independent clinical trial
[34]	Objective: Detection of leukemia and its types Data set: 220 blood smear images from healthy individuals and patients with leukemia Methodology: support vector machine	Performance: Accuracy: Above 80% Strengths: Use of 3 segmentation methods Broader range of leukemia classification (types and subtypes) Limitations: Costly method based on imaging data Validation: Internal validation (train test split)
[122]	Objective: Automated detection of malignant lymphoma Data set: Prepared histopathologic images (388 sections, 259 diffuse large B-cell lymphomas, 89 follicular lymphomas, and 40 reactive lymphoid hyperplasia) Methodology: Deep neural network classifier	Performance: Accuracy: 97% Strengths: High accuracy outperforming 7 pathologists Model ensemble comprising 3 classifiers Limitations: Classifier requires a manual annotation Model not able to classify all the subtypes Validation: K-fold cross validation repeated 5 times
[37]	Objective: Multiclassification of leukemia Data set: 100 blood smear images Methodology: Neural network classifiers	Performance: Accuracy: 97.7% Strengths: Two-step neural network classifier Limitations: Limited data set (100 blood smear images) Validation: Internal validation (90 images used for training and 10 kept for validation)
[133]	Objective: Leukemia and lymphoma diagnosis Data set: 283 blood and bone marrow sample images from patients with leukemia and lymphoma Methodology: Decision tree	Performance: Correctness: 95% Strengths: Application of the LASSO algorithm for regularization Model robustness and strength against false negatives Limitations: Complexity of the decision tree and the risk of overfitting through the production of too large trees Validation: 30-fold cross validation
[66]	Objective: Leukemia image segmentation Data set: The Acute Lymphoblastic Leukemia Image Database Methodology: HSCRKMb/particle swarm optimization/K-means	Performance: Accuracy: 80% and above Strengths: Use of 7 machine learning methods Application of soft covering rough approximation Limitations: Suitable for medical images only Application on multiple color images increases the processing time Validation: Different train/test sizes were used for model evaluation
[144]	Objective: Determining the most predictive features for acute lymphoblastic leukemia identification Data set: 94 pediatric patient samples collected from the Department of Hematology and Oncology, Children Hospital and Institute of Child Health, Lahore Methodology: Random forest, boosting machine, C5.0 decision tree, and classification and regression trees	Performance: Accuracy: 87.4% Strengths: High accuracy Balanced data set Limitations: Small-scale study Few machine learning models Socioeconomic risk factors not selected automatically Validation: Internal validation (train/validation data) 10-fold cross validation
[31]	Objective: Leukemia diagnosis and its subtypes Data set: 200 blood smear images extracted from Vidyalankar Institute of Technology, Mumbai and online databases Methodology: support vector machine	Performance: Accuracy: 97.8% Strengths: Good detection accuracy Thorough image segmentation process Limitations: Challenging detection process due to the irregularity of the cancer cell’s shape and nucleus Use of only support vector machine for classification

^aCNN-SOM: convolutional neural network-self-organizing map.

^bHSCRKM: histogram-based soft covering rough K-means clustering.