Acute lymphoblastic leukemia (ALL) is an extremely severe hematological cancer necessitating accurate detection and classification for optimal treatment. Conventional diagnostic techniques, however effective, may encounter constraints in precision and efficacy. Recent breakthroughs in machine learning (ML) and deep learning (DL) technology present interesting possibilities for improving diagnostic procedures. The incorporation, of these computational methods into clinical practice is inconsistent, exhibiting disparities in the performance of models, quality of data, and interpretability among research. This systematic review evaluates the landscape of machine learning and deep learning methodologies for detecting and classifying ALL. It will assess the efficacy, resilience, and therapeutic relevance of these procedures in comparison [21,22] and [23], to traditional methods. The review will also pinpoint current problems, including data diversity, validating models, and the necessity for explainable AI in healthcare environments. This study aims to emphasize the potential of machine learning and deep learning in enhancing diagnostic precision and effectiveness in acute lymphoblastic leukemia, thereby informing future studies and clinical applications in hematology.

The selection of effective segmentation techniques is important in distinguishing the area of interest [26]. The segmented images are shown in Fig. 4 and deep learning-based segmentation framework is shown in Fig. 5. The literature different segmentation algorithms were used to achieve the goal.

Ko et al. [82] proposed a method for the classification of WBC using random forests and decision trees with effective results. Ramoser et al. [83] came up with a technique for the identification and classification of lymphocytes using an SVM into its subtype, [84,85]. Tai et al. [86] suggested a method that is used to segment and classify various components in blood slide images to separate the nucleus and cytoplasm. Geometric features with multi-class SVM are used to classify them. Mohapatra et al. [87] used features like the Hausdroff Dimension with contour signature to detect nuclei boundaries and then classify them using SVM. Studies [88,89], Rawat et al. outlined a system to distinguish between normal and ALL leukocytes using Gray-level Co-occurrence matrix GLCM and shape features, SVM, and random forest for classification respectively. It gives 86.7% accuracy for nuclei, 72.4% for cytoplasm, and overall accuracy of 89.8%. MoradiAmin et al. [90] gives a framework that can categorize leukemia into its four types using SVM. Pan with others in [91] used the mean shift technique for the partition of the nucleus of WBCs, using the learning-by-training method SVM. Mohapatra et al. [92] came up with a color-based clustering technique using k-means and Fuzzy Possibilistic C-means in combination with Gustafson Kessel for the segmentation nuclei from white blood cells to identify ALL Blasts using the SVM classifier. James et al. [93] segment WBC to use the k-means clustering approach to find AML, normal, and cancerous cells were classified using SVM. Asadi et al. [94] used a holographic method for the detection and categorization of leukemic cells using Zernike moments for feature extraction and then classified the cells with k-nearest neighbor (KNN) using minimum mean distance. Authors in [95,96] proposed a method that identifies ALL using CMYK color format and then the Zack threshold method, classified the cells using KNN, and gives good accuracy. Di Ruberto with others in [97] came up with a framework consisting of the combination of KNN and SVM that can segment and classify different components of blood and their nucleus and cytoplasm, resulting in an accuracy of 99% using the ALL-IDB data set.

Sahlol et al.’s research [98] presented a framework based on CNN with a visual geometry group (VGG) net model. They enhanced the slap swarm algorithm statistically to classify the subtypes of leukocytes. Rehman et al. [99] segmented the ALL blasts using the threshold method and then used the CNN-based Alexnet model which gives efficient results and gives 97.78% accuracy. Shafique et al. [100] suggested a framework composed of CNN with an Alexnet model for identifying ALL using the ALL(IDB) database and gave the result of 96.06% accuracy. Loey et al. [101] proposed a model for the diagnosis of ALL, CNN with Alexnet model is used on the ALL (IDB) database and classifies images into two classes, i.e., normal and affected giving 100% results. Mallick PK with others in [102] came up with a framework based on Deep Neural Network (DNN) through which they classify two classes, i.e., ALL and AML, and resulted in 98.2% accuracy. In [103], the research work develops a diagnosis and detection of ALL that enhances different blood images with adaptive sharpening and then uses deep learning techniques [104]. A thorough overview of the deep learning architectures in Fig. 6, including model architecture, training procedure, and hyperparameter values, is required to comprehend and replicate the work fully. This includes a comprehensive description of the model’s general architecture as shown in Fig. 6, including the very first step of dataset collection and pre-processing of that data, after this the segmentation algorithms are selected and the region of interest is segmented as shown in Fig. 5 after this step robust classification technique is selected and layout of layers are set according to the need of the required results (e.g., convolution, recurrent, fully connected), functions for activation, and any other relevant components. The training method is thoroughly detailed, including information about the datasets utilized, the number of epochs, batch sizes, and the optimization techniques (such as Adam or Stochastic gradient descent (SGD)) used to achieve the best results [105]. Furthermore, hyperparameter settings should be properly specified, such as learning rates, dropout rates, weight initialization methods, and regularization approaches. Providing these facts encourages transparency, enabling others to duplicate the study and evaluate the effectiveness of the model under identical conditions, which is critical for scientific accuracy and field advancement [103].

A Convolutional Neural Network (CNN) architecture can be built with many essential components and hyperparameters that influence the model’s functionality. This article provides a summary of the standard layers and parameters utilized in the construction of a CNN, together with their traditional values [105].

In recent years, different CNN models have been used to those results in more accurate and efficient ways [107] as given in Table 3. Graphical representations of different machine and deep learning techniques were given in Fig. 7. YOLOv3 [108] and YOLOv4 [109] are some of the recent faster CNN models, these are models with high computational efficiency for localizing the region of interest and classification tasks. Ai-Qudah et al. [110] used YOLOv2 [111] to classify ALL into normal/healthy cells efficiently. Khandekar et al. [112] suggested a classification technique based on YOLOv4, classifying healthy and ALL-blasts efficiently. Authors in [50] also proposed medical robotic diagnosis based on deep learning methods. Duggal et al. [113] suggested CNN CNN-based framework for cancer detection with a deconvolution layer that can convert the images to Optical Density. It also involves back-propagation to de-convolve the images to tissue-specific for the next layer as input. The authors suggested a unique technique based on the examination of the available white blood cells (WBC) to identify acute lymphoblastic leukemia (ALL) in the blood’s peripheral circulation. In contrast to previous approaches described in the literature, this technique combines a histopathological transfer learning process with a lightweight CNN. This is achieved by introducing a CNN with less learnable parameters that mimic Local Binary Patterns (LBP) and learning to recognize different types of histology tissues. It then fine-tunes this CNN on the ALL database to categorize each cell as either normal or lymphoblast. Sulaiman et al. [114] suggested a hybrid model with an accuracy claim of an F1 score of 0.929 for the classification of ALL into healthy and infect groups, based on ResNet and SVM. Reference [115] suggested the automated identification of healthy cells and ALL to compensate for the manual analysis deficiencies of an expert. The model utilized is YOLOv3, which generates low loss values and elevated mAP evaluation values through a transfer learning technique.

The evaluation’s findings show that the YOLOv3 model can distinguish between ALL and healthy cells. In previous research [116], acute lymphoblastic leukemia has been classified using Mask R-CNN on microscopic pictures of white blood cells, which can effectively and efficiently support the diagnosing process. Menagadevi et al. [117] proposed a technique that can predict ALL, they applied k-means clustering for segmentation with CNN and resulted in an accuracy of 98%, specificity of 97%, and sensitivity of 98.2%. Some of the frameworks that perform with high accuracy are given in Figs. 8–11. Although machine learning strategies present benefits regarding interpretability and reduced processing requirements, deep learning methods deliver higher precision and feature extraction abilities. A thorough evaluation of these factors like accuracy, adaptability, computational expense, and interpretability, will yield a full insight into their respective functions in the recognition and classification of ALL. Using huge dataset information, pre-trained deep learning algorithms can be fine-tuned for particular uses with small datasets [101]. This helps in medical applications with minimal labeled data.

Diverse models such as CNNs, VGGNet, and YOLO demonstrate differing efficacy contingent upon dataset size and image noise levels. Convolutional Neural Networks are versatile and can adjust to diverse tasks, although they might require greater datasets to generalize proficiently [112]. VGGNet, recognized for its deep design, performs exceptionally well with extensive, clean datasets owing to its capacity to capture complex features; yet it is prone to overfitting on smaller datasets that demand significant processing resources. YOLO, developed for real-time object recognition, excels with extensive datasets, effectively identifying things in noisy environments; nevertheless, its accuracy may diminish if the training data is not diverse [118]. In conclusion, although CNNs provide adaptability, VGGNet excels in meticulously organized settings, and YOLO, achieves a compromise, between speed and efficacy in variable contexts, rendering the selection contingent upon particular applications and data attributes. Deep learning in clinical environments where the accuracy, of data may fluctuate due to varying processing of samples or imaging conditions.

This framework [99] aims to enhance the diagnosis of Acute Lymphoblastic Leukemia (ALL) with a computer-assisted approach that integrates image processing and deep learning methodologies. The suggested method entails categorizing ALL into its subtypes and normal reactive bone marrow utilizing stained bone marrow images as shown in Fig. 8. The model was trained in bone marrow pictures using robust segmentation and deep learning with a convolutional neural network (CNN) to attain precise classification outcomes. The experimental findings demonstrated that the suggested method surpassed conventional classifiers, including Naïve Bayes, KNN, and SVM, attaining a remarkable accuracy of 97.78%. This method provides a significant resource for pathologists, improving the precision and efficacy of ALL diagnoses. Reference [121] shown in Fig. 9 introduces an effective deep Convolutional Neural Network (CNN) framework for the automated diagnosis of Acute Lymphoblastic Leukemia (ALL), tackling the issue of necessitating extensive datasets for training. The suggested approach integrates depthwise separable convolutions, linear bottleneck architecture, inverted residuals, and skip connections, in conjunction with an innovative probability-based weight factor to amalgamate MobileNetV2 and ResNet18. The methodology, corroborated on the ALLIDB1 and ALLIDB2 benchmark datasets, attains exceptional accuracy—99.39% and 97.18% for 70% training and 30% testing, and 97.92% and 96.00% for 50% training and testing. It surpasses contemporary transfer learning methodologies regarding sensitivity, specificity, accuracy, precision, F1 score, and ROC. This framework [122] for the application of Multi-Attention EfficientNetV2S and EfficientNetB3 deep learning architectures, optimized using transfer learning, to differentiate between normal and blast cells in microscopic blood smear images for the identification of acute lymphoblastic leukemia (ALL) shown in Fig. 10. The Multi-Attention Mechanism decreases model complexity and enhances generalization by altering the last blocks of both models and incorporating supplementary layers. The proposed models demonstrated exceptional accuracy, with EfficientNetV2S reaching 99.73% and EfficientNetB3 achieving 99.25%. The proposed methodology surpassed previous methods, exhibiting enhanced efficiency in leukemia identification compared to other models. Fig. 11 shows a non-invasive, CNN-based methodology for the diagnosis of Acute Lymphoblastic Leukemia (ALL) utilizing medical imaging [137]. The model integrates an Efficient Channel Attention (ECA) component with VGG16 to optimize feature extraction, hence enhancing the categorization of malignant and normal cells. The method employs data augmentation to enhance both the quality and quantity of training data, while simultaneously mitigating subject-level variability by partitioning the dataset into seven folds. The proposed model attained an accuracy of 91.1%, indicating its capability to aid pathologists in identifying ALL.

The ALL classification technique utilizing the MobileNetV2-SVM architecture proposed by Das et al. [107] attains the highest accuracy (98.21%) and the optimal F1 score (0.9828). It provides commendable performance owing to the synergistic advantages of MobileNetV2-based extraction of features and SVM-based classification. The majority of research concentrates on identifying ALL by categorizing individuals as either healthy or affected by ALL, whereas only a limited number of studies prioritize the further classification of ALL into its subtypes (L1, L2, and L3). Table 4 illustrates the classification performance of AlexNet, as proposed in [101], which categorizes white blood cells into healthy and three subtypes of acute lymphoblastic leukemia: L1, L2, and L3. L1 is identified as the most properly defined subtype within the group. The proposed approach attains an overall accuracy of 97.78%. The morphological resemblance in ALL and healthy images, the imbalanced dataset, and the existence of intersubject variability may compel a system to acquire subject-specific features instead of class-specific features [138]. Consequently, these characteristics complicate the ALL classification.

Table 5 demonstrates that the majority of research is predicated on transfer learning methodologies, based on this table data, owing to their capacity to deliver favorable outcomes with limited datasets. Table 6 represents different publicly available datasets [140], in which ALLIDB1 and 2 are standard and the most popular datasets. Images examples are given in Figs. 12 and 13 while a graphical representation of the usage of datasets is given in Fig. 14. Results of different methods using these datasets were also shown graphically in Fig. 15 and ROC Curve in Fig. 16. The diversity and representativeness of datasets such as ALLIDB1, ALLIDB2, BCCD, ATLAS, and C-NMC are essential for the development of effective models for the diagnosis of acute lymphoblastic leukemia (ALL). ALLIDB1 and ALLIDB2 [141] offer images primarily related to leukemia; nevertheless, their representativeness may be constrained by the demographic and diagnostic attributes of the included individuals, thus affecting the model’s adaptability to other samples. The BCCD dataset, although useful for general blood cell evaluation, may not adequately represent the distinctive morphological characteristics of ALL cells. ATLAS provides a comprehensive framework including various blood illnesses, hence augmenting its diversity; yet it may still be deficient in comprehensive illustrations of all subtypes. C-NMC emphasizes cell morphology, essential for leukemia identification, although may not account for the heterogeneity in staining and imaging circumstances present in actual clinical environments. Therefore, although these datasets are valuable resources, their limitations in variety and representativeness must be recognized to guarantee that models trained on data can be applied effectively to diverse patients.

The most current developments in the field are considered as this article briefly studies deep learning and conventional machine learning techniques. Classification using traditional machine learning requires additional techniques for accurately segmenting unshaped and overlapped cells by extracting features like texture, color, and geometrical features and then normalizing these using different normalization algorithms. While in deep learning all the phases mentioned above were included in a single algorithm. Deep learning models, especially convolutional neural networks (CNNs), may autonomously discern pertinent features from unprocessed data, thereby diminishing the necessity for manual feature engineering [135]. CNNs are adaptable and essential for many imaging tasks, however noise and overfitting must be managed. VGGNet might benefit via transfer learning and excels on large datasets but is highly computational and sensitive to small amounts of data without fine-tuning. In real-time object detection, YOLO works well with large datasets but may struggle with short datasets or excessive noise [108].

Although intended for real-time detection, it demands substantial resources, particularly for training on intricate datasets; yet, its architecture facilitates expedited inference, rendering it appropriate for applications where speed is paramount. The selection of a model must equilibrate performance, resource availability, and the particular requirements of the application, especially in contexts with constrained computational capability [101]. Deep learning techniques demand extensive processing resources, such as fast GPUs or TPUs, as well as plenty of RAM and storage to handle massive datasets and complicated models. Cloud computing platforms frequently offer scalable ways to satisfy these objectives effectively.

This is especially advantageous with intricate datasets, such as genetic or imaging data. However, in the deep learning approach, large-scale datasets are required to train the model efficiently and result in classification. Managing potential biases in the utilized datasets, including class imbalances, is essential for a thorough assessment of both conventional machine learning (ML) and deep learning techniques [86]. An imbalanced dataset can substantially impact the performance of models, as algorithms may exhibit bias towards the dominant class, resulting in deceptive accuracy numbers [142]. For example, if a dataset has a significant predominance of images from a certain category and a scarcity from a different one a model may attain elevated accuracy merely by identifying the majority class, but poorly generalizing to the minority class.

Deep learning algorithms, however effective in recognizing patterns from intricate data such as images, genomic, and flow cytometry data, face certain obstacles. A primary restriction is the necessity for extensive, high-quality datasets with annotations for training, which are sometimes limited, especially for rare diseases such as ALL. The models may encounter difficulties in generalizing across various patient populations or clinical environments, resulting in diminished accuracy in varied or underrepresented groups [140]. Moreover, deep learning models frequently operate as “black boxes,” complicating the interpretation of the rationale behind predictions, which is crucial in medical environments where explainability is vital for clinical decision-making. Overfitting to trained data, particularly when sample sizes are limited or unrepresentative of the larger population, can pose a considerable issue. Moreover, deep learning algorithms may exhibit sensitivity to noise and abnormalities in medical imagery or data, potentially resulting in inaccurate diagnoses [90]. A comprehensive examination of these limitations, coupled with potential solutions like data augmentation, transfer learning, and explainable AI techniques, would enhance the understanding of the challenges related to the application of deep learning in ALL diagnoses and inform future advancements in this domain [135].

To conduct a more comprehensive analysis, it is essential to examine how each study addressed class imbalance. Methods such as exceeding the minority class, reducing the majority class, or utilizing synthetic data generation techniques (such as Synthetic Minority Oversampling Technique SMOTE) help alleviate this problem. Moreover, employing performance criteria that account for class distribution, like precision, recall, F1 score, or area under the receiver’s operating characteristic curve (ROC) as shown in Fig. 16, provides a more equitable assessment of model efficacy [84]. This study will elucidate the merits and shortcomings of classical machine learning and deep learning algorithms while offering insights into the influence of dataset features on their performance, so facilitating a more comprehensive comparison.

To verify the research work done in processing medical images, different standard datasets are freely available for several types of blood diseases as mentioned in Table 4. Nowadays researchers have put some effort into analyzing data and finding new frameworks using already existing datasets and applying their methodologies using transfer learning techniques in deep learning algorithms as shown in Table 3 and show some good results, as they used already existing ideal data to investigate the issue. It is not easy to compare the accuracies using one dataset and the same algorithms. Most importantly this process needs to be verified by pathologists to save more human lives. The framework should be made with more care and use datasets having a large number of diseases. Which is one of the most difficult tasks. ALL detection and classification can face different challenges:

The presence of noise and inhomogeneity, weak edges, and overlapping cells can affect the results of segmentation.