With the clear benefits for patients and medical professionals, the importance of computer-aided diagnosis (CAD) has suddenly become obvious in modern healthcare. Also, the automatization of processes conventionally maintained by human professionals becomes increasingly important [1]. The image analysis process can be supported using networks that perform multilayer analysis of patterns, widely known as artificial intelligence (AI). Computer networks can exhibit the result with lower error bias if supported by large datasets of input data [2]. Medical imaging focused on pattern detection is usually supported by AI algorithm, and it is used as a powerful aid in the three most important steps of decision-making in the medical imaging task: recognition (assignment to the class) [3], detection (image segmentation), and result description (conversion of the results into natural language). Implementing an AI algorithm can contribute to standardizing the diagnosis process and significantly decrease the time required to diagnose pathology and describe the result [4].

Leukemia is a lethal cancer-related disease that significantly causes higher mortality rates worldwide and affects all age groups, i.e., adults and children [5]. Specifically, it is linked with White Blood Cells (WBC), which can cause damage to the blood or bone marrow and increase immature lymphocytes. The correct classification of WBC is critical to find the disease’s nature [6]. They are classified based on the composition of the cytoplasm. Most classifications of WBC are Lymphocytes and syndromes resulting in Acute Lymphoblastic Leukemia (ALL). In General, leukemia can be classified into two types: chronic and acute leukemia. Without any specific medication, the total recovery time of acute leukemia was three months, whereas the recovery time of chronic leukemia was found to be greater than that of acute leukemia [7]. ALL can be referred to as one common type of acute leukemia accountable for nearly 25% of every childhood tumor.

In the primary step, it seems to be in the bone marrow and can be later dispersed all over the human body [8]. In a healthy person, the WBC development is based on body necessities, but in terms of leukemia, it can be formed abnormally and turns out to be ineffective. Hence, a reliable and rapid cancer diagnosis becomes a severe need for fruitful treatment to increase survival rates. At present, a manual examination of blood samples gained with the help of microscopic images was made to detect this disease, which can be less accurate, often very slow, and time-consuming [9]. Also, in the microscopic study, the shape and appearance of leukemic cells are like normal cells, which makes diagnosis challenging [10]. Convolutional Neural Networks (CNNs) have been used in Deep Learning (DL) to classify images for years, but there is still a gap in enhancing their performance, efficiency, and learning process.

The MPADL-LCC algorithm was applied to classify medical images based on the Marine Predator Algorithm with Deep Learning (MPADL-LCC). The projected MPADL-LCC system utilizes the bilateral filtering (BF) technique as a preprocessing stage. In addition, the MPADL-LCC technique employs Faster SqueezeNet for feature extraction with the Marine Predator Algorithm (MPA) as a hyperparameter optimizer. Eventually, the denoising autoencoder (DAE) methodology can be executed to detect and classify leukemia cancer accurately. The series of simulations of the MPADL-LCC approach can be tested on the medical imaging dataset.

Devi et al. [11] presented an innovative technique called the Gaussian Blurring Hue Saturation Value GBHSV-Leuk approach for classifying and segmenting ALL cancer cells. The first stage includes preprocessing and blurring the reflections and noise in the image using the Gaussian Blurring (GB) method. The next stage contains segmentation through the morphological operations and Hue Saturation Value (HSV) technique for distinguishing the background and foreground colors, which enhances the prediction accuracy. Jawahar et al. [12] offered a deep neural network (DNN)-based (ALNett) technique to classify microscopic WBC images, and depth-wise convolutions having different dilation rates were used. For extracting robust global and local features out of microscopic images for precise forecasting of ALL, the cluster layers with convolution, max-pooling, and normalizing procedures are utilized to offer enriched contextual and structural details. Ramaneswaran et al. [13] devised a hybrid Inceptionv3 and XGBoost methods for classifying ALL from microscopic WBC images. In this technique, the XGBoost method performs as classification, and head Inceptionv3 performs as image feature extraction.

Sunil et al. [14] introduced a completely automatic ALL detection technique that depends on DNN. ALL refers to a condition that affects the leukocytes. Children are likely to suffer from melanoma. Unlike other cancers, it does not create tumors, which is the main difficulty faced with this type of cancer and makes it very difficult to detect. Manual microscopic testing procedures have been employed before automation, but they are error-prone and time-consuming. However, this study deals with medical data and needs better accuracy and efficiency; hence, this system used various CNNS architectures. The research presents a potential deep CNN structure for mitigating this problem and yielding more precise ALL detection. Prostate cancer (PRC) stands as a predominant global cause of mortality. The imperative for enhancing healthcare services lies in the early identification and characterization of the PRC. Utilizing microarray gene expression data (GED), a contemporary approach intertwining DL and machine learning (ML) methodologies, fortified by multiple optimization tools, presents a promising avenue for the effective categorization of PRC [15].

In this technique, a new probability-related weight factor was recommended, which has a vital role in hybridizing professionally Mobilenet-V2 and ResNet-18 while protecting the merits of both methodologies. Zakir Ullah et al. [16] described a non-invasive, CNN-related technique that sizes medical images to make tasks. The presented solution, an attention module termed Efficient Channel Attention (ECA) comprising CNN-related techniques, utilizes VGG16 to derive better-quality deep attributes and image data, leading to superior classifier outcomes and feature representations. A technique with an ECA module will help solve morphological similarity among healthy cells and ALL cancer images. In [17], a potential CAD system for ALL detection was devised. This technique covers two stages. In the first phase, the WBC was segmented in microscopic blood images. The second stage extracts significant features, such as texture and shape features, from segmented cells. Eventually, K-nearest neighbor (KNN) and Naive Bayes (NB) methods are applied to find the normal and abnormal segmented cells on the extracted features. The histological diagnosis of these cancers is critical in choosing the best treatment option. On either front, early and accurate diagnosis reduces the probability of death.

This study employs an innovative deep-learning model inspired by marine predators, MPADL-LC3, to classify lung and colon cancer. As cancer spreads from one part of the body to another, metastasis is a complex process [18]. Through a comparative analysis, the study exhibits the superior performance of the MPADL-LC3 system across various parameters. The study introduces an advanced hybrid methodology for classifying WBC subtypes. Initially, to extract optimal deep features, refined and segmented WBC images undergo transfer learning on established neural networks, including DenseNet201 and Darknet53. Subsequently, the feature vector undergoes filtration using an entropy-controlled marine predator algorithm (ECMPA). This meta-heuristic optimization technique, inspired by nature, selectively retains robust features and discards weaker ones. The resultant reduced feature vector is inputted into various baseline classifiers with diverse kernel values to achieve effective subtype classification. The efficacy of the proposed approach is demonstrated through validation on a publicly available dataset, encompassing 5000 synthetic images representing five distinct WBC subtypes [19].

The marine predator algorithm is a meta-heuristic optimization method rooted in the survival-of-the-fittest concept. However, the conventional marine predator algorithm exhibits a unidirectional search approach, limiting its ability to converge effectively and often leading to stagnation at local minima. An innovative opposition-based learning feature has been introduced to enhance exploration capabilities and facilitate productive convergence towards global optima [20]. Although the automated classification of White Blood Cells (WBCs) using deep convolutional neural networks indicates promising accuracy, it grapples with substantial processing costs attributed to extensive feature sets. Through dimensionality reduction, intelligent feature selection emerges as a critical strategy to enhance model performance while mitigating computational complexity [21]. Peripheral Blood Smear (PBS) images are frequently favored for initial malignancy screenings in ALL diagnoses. Manual perusal of PBS images in laboratories for cancer detection is susceptible to errors due to interoperability issues and human fatigue. In addition, the non-specific nature of symptoms in blood smear images and their intricate morphological structures pose significant challenges when using trained data. Advanced algorithms outperform traditional methods, offering effective solutions to image processing challenges [21]. A novel CNN-based framework was introduced to improve accuracy. The proposed CNN model was tested on the primary biliary cholangitis (PBC) dataset normal DIB (Direct Immunoblotting). The results of the studies showed that the CNN-based framework for blood cell categorization achieves a 99.91% accuracy on the PBC dataset. When compared to previous results given in the literature, the proposed convolutional neural network model performs competitively [22,23].

This research aims to establish an innovative framework for predicting lung cancer using Electronic Claims Records (ECRs) through a meticulously designed exclusion process, employing a state-of-the-art sequence-based transformer. It unveils pioneering outcomes by applying a disease prediction model to the entire population of Taiwan. It provides comprehensive evaluations of lung cancer prediction, presenting the proposed methodology as an advanced framework for predicting disease risk based on claims data. Although clinical tests and imaging methods are conventionally employed for diagnosing metastasis, they are typically administered after the event. Metastasis prediction or early diagnosis is critical for treatment planning since it affects patient survival. The method created in this study will be valuable in predicting metastasis and, as a result, in choosing therapy options for cancer patients [24,25].

This study presents a diverse array of conventional Machine Learning and cutting-edge deep learning algorithms designed for the intricate classification of driver somatic non-synonymous mutations. Using a 70-feature annotation derived from medical and statistical tools, the methodology, Driver-Oriented Genomics Analysis (DrOGA), surpasses individual and ensemble methods based on the extensive analysis of a comprehensive database from various public data sources [26]. The developed computer-aided detection technique discriminates lung tumors and extracts tumor-related attributes from chest computed tomography images. After the feature selection process, a classifier is applied based on the Naive Bayesian network, utilizing eight selected image features. In addition, a multilayer perceptron classifier is employed that integrates two genomic biomarkers. This dual-classifier approach enhances the system’s robustness, incorporating visual and genomic information for more accurate and comprehensive leukemia detection 22. This system accurately predicts the risk of cancer recurrence. Training and evaluation involve a dataset comprising 79 stage I Non-small cell lung cancer (NSCLC) patients, utilizing a synthetic minority oversampling strategy and a leave-one-case-out validation approach. In addition, a fusion method integrates prediction scores from the two classifiers for enhanced accuracy [27].

A multi-class classification problem is created by binning the objective variable into four bins to accurately forecast the patients’ survival months. Using a multi-tier categorization system improves the accuracy of the results. Although the final numerical results are biologically significant, it is advised that a therapeutically meaningful model be derived with caution [28]. This research investigates numerous aspects of machine learning and deep learning approaches, focusing on feature selection and categorization. The reviews and comparisons demonstrate the parameter algorithm’s resultant correctness in the feature selection and classification process by assessing the testing outcomes in different datasets. In addition, different methodologies are used in this study to focus on the difficulties and identification of variables [29]. Exploring both Acute Myeloid Leukemia (AML) and ALL, this investigation utilizes gene expression data from a diverse cohort to predict leukemia subtypes using three distinct machine learning algorithms: XGBoost, Random Forest Classification, and Artificial Neural Networks. Before applying the algorithms, the dataset underwent dimensionality reduction through principal component analysis (PCA) [30]. Focusing on binary categorization, the primary goal of this study is to optimize the efficiency and accuracy of classifying leukemia samples into Acute Lymphocytic Leukemia and Acute Myelocytic Leukemia. Various feature selection approaches were employed and compared, ultimately selecting the dimensionality reduction algorithm that demonstrated the highest classification accuracy with the fewest genes (features). Computer science researchers have developed numerous automatic approaches for selecting algorithms and/or hyper-parameter values for a particular supervised machine learning problem, making machine learning accessible to layperson users with minimum computing skills [31]. Hyper-parameter tuning is crucial in determining the ideal machine learning parameters [32]. This indicates that the hyper-parameter values for certain leukemia traits were successfully determined using the Hybrid Optimization of Resource Distributions (HORD) technique. The HORD approach also improved the model’s performance by adjusting and locating the ideal set of hyper-parameters for the CNN model [33]. This model maximizes feature selection and refinement from high-dimensional data by fusing machine learning with image processing [34].

This study formulated a novel MPADL-LCC system for detecting and classifying leukemia cancer on microscopic images. The projected MPADL-LCC methodology encompasses BF-based noise removal, Faster SqueezeNet, MPA-based hyperparameter optimizer, and DAE-based classification. Fig. 1 illustrates the overall flow of MPADL-LCC methods. Integrating the Marine Predators Algorithm with deep learning provides a unique approach to leukemia detection. It efficiently extracts features from complex medical images using the algorithm’s strengths, which is crucial for accurate cancer classification. Deep learning enables higher accuracy in distinguishing leukemia from non-leukemia cells. Scalable and adaptable, the model can handle large datasets and improve with additional data, enhancing its performance over time.

This study employed a robust computing environment with powerful CPU and GPU capabilities to manage the intensive computational tasks required for training neural networks to implement the leukemia cancer classification system using deep learning. It utilized popular frameworks such as TensorFlow, PyTorch, or Keras, ensuring all necessary libraries were installed. This study designed a convolutional neural network architecture for medical image analysis, specifying layer types (e.g., convolutional, pooling), activation functions, and layer depths based on previous research. A labeled dataset of medical images containing leukemia and non-leukemia cases was prepared for supervised learning. Then, an optimizer such as Adam or RMSprop was employed to minimize training loss. For evaluation, key evaluation metrics (accuracy, precision, recall, F1 score, confusion matrix) were defined for model assessment. The dataset was split into training, validation, and test subsets (e.g., 70-15-15 ratio) to train and validate the model while preventing overfitting. The hyperparameters are fine-tuned based on validation performance, and the final model was evaluated on the test set. In addition, the developed model was constructed using Python in conjunction with the Keras deep learning framework, and all computations are executed on Google Colab utilizing a 12GB NVIDIA Tesla K80 GPU. The experimental configuration encompasses the dataset used, and the model’s performance was rigorously validated. The simulation result of the MPADL-LCC algorithm was tested on the C-NMC_Leukemia dataset available from [16].

The Marine Predators Algorithm aids in efficient feature extraction from medical images inspired by marine predators’ foraging behavior to capture relevant leukemia patterns. The model achieves enhanced classification accuracy by learning intricate image patterns by integrating deep learning, particularly CNNs. This combination results in a robust classifier capable of handling variations in image data quality and resolution. Improved accuracy promises early leukemia detection, which is crucial for timely intervention. These advancements have significant clinical implications, potentially improving diagnostic processes and patient care. Future research aims to refine the algorithm, optimize parameters, and validate its performance across diverse clinical settings.

Performance measures are described in Eqs. (22)–(25), and they are accuracy, recall, precision, and F1 score. True Positive (TP) denotes the number of correctly detected positive samples; True Negative (TN) is the number of correctly classified negative samples; False Positive (FP) is the number of falsely detected positive samples; and False Negative (FN) is the number of undetected positive samples. (22)Accuracy=No. of correct predictionsTotal no. of predictions (23)Precision=TPTP+FP (24)Recall=TPTP+FN (25)F1−Score=2∗Precision∗RecallPrecision+Recall

(26)Confivence Interval=Estimated metric±z×Estimated metric×(1−Estimated metric)nwhere:

Estimated MetricEstimated Metric is the observed performance metric (e.g., accuracy, sensitivity, and others),

In order to assess the robustness of the hybrid model, it is necessary to test it on datasets that differ in picture quality, resolution, and illness subtypes, reflecting the variety present in clinical practice. This method guarantees that the model can adjust to various imaging situations. Standard metrics, including accuracy, precision, recall, and F1 score, evaluate the model’s performance across these many datasets. This allows for understanding the model’s generalization and consistent performance in real-world circumstances. Extensive testing on various leukemia picture sets is necessary to confirm the model’s applicability in medicine by validating its dependability and possible clinical impact under different imaging settings. It can evaluate if the model’s performance holds steady over various groups of photos by dividing the dataset into folds and training/testing the model on various combinations.

This study used the dataset_2 for the second validation phase. In addition, this dataset has insufficient microscopic blood samples to adequately train the proposed model. In order to overcome this restriction, data augmentation techniques were employed to improve the diversity of photos in the training set. The training process was improved by adding several image changes to the dataset. The expanded train set was then utilized to train the model. This validation phase’s results, which involved randomly rearranging the train and test sets, are encouraging. In particular, an accuracy of 98% was attained in the second run of testing with the dataset_2. Table 1 describes the Samples for testing and training data. Fig. 3 describes the confusion matrics of the leukemia data for dataset_2. The database contains 530 images, with 360 cancer images and 170 healthy images, as listed in Table 2. Fig. 4 illustrates the sample images of the system.

In Fig. 5, the confusion matrices of the MPADL-LCC methodology are inspected under varying numbers of epochs. The MPADL-LCC algorithm perfectly recognizes cancer and healthy images.

In Table 3, the overall classifier outcome of the MPADL-LCC approach is inspected under several epochs. Fig. 6 depicts the sensy and specy results of the MPADL-LCC technique. The outcomes showed that the MPADL-LCC methodology recognizes cancer and healthy classes under each epoch. For instance, with 500 epochs, the MPADL-LCC technique achieves average sensy and specy of 95.56% and 95.56%, respectively. In addition, with 1500 epochs, the MPADL-LCC approach achieves an average sensy and specy of 95.84% and 95.84% correspondingly. In addition, with 3000 epochs, the MPADL-LCC method achieves an average sensy and specy of 98.87% and 98.87%, correspondingly. The model has a 99% confidence level: z = 2.576.

Fig. 7 depicts the accubal, precn, and Fscore outcomes of the MPADL-LCC methodology. The outcomes specify that the MPADL-LCC algorithm recognizes cancer and healthy classes under each epoch. For example, with 500 epochs, the MPADL-LCC method achieves average accubal, precn, and Fscore of 95.56%, 94.63%, and 95.07% correspondingly. Also, with 1500 epochs, the MPADL-LCC method achieves an average accubal, precn, and Fscore of 95.84%, 95.16%, and 95.49%, correspondingly. With 3000 epochs, the MPADL-LCC method achieves an average accubal, precn, and Fscore of 98.87%, 98.14%, and 98.50%, correspondingly.

TRAC (Training Accuracy) and VDAC (Validation Accuracy) of the MPADL-LCC method are studied in the Leukemia cancer solution in Fig. 8. The outcome stated that the MPADL-LCC approach has maximal results with enhanced values of TRAC and VDAC. Visibly, the MPADL-LCC method has greater TRAC outcomes.

The TRLS (Training Loss) and VDLS (Validation Loss) of the MPADL-LCC methodology are tested on Leukemia cancer performance in Fig. 9. The outcome inferred that the MPADL-LCC algorithm has a greater outcome with the lesser values of TRLS and VDLS. The MPADL-LCC system has minimal VDLS performance.

An evident PR curve of the MPADL-LCC algorithm on the test database is presented in Fig. 10. The outcome highlighted that the MPADL-LCC methodology has greater PR values in two classes.

A comprehensive ROC curve of the MPADL-LCC method on the test database is defined in Fig. 11. The result indicated that the MPADL-LCC approach can classify two classes.

Table 4 and Fig. 12 show a brief comparative outcome of the MPADL-LCC algorithm with recent DL methods [39]. The outcomes inferred that the YOLOv4, EfficientNet-B2, and Inception-ResNetV2 algorithms demonstrate moderately closer classification outcomes. Then, the Inception V3, Xception-ensemble, and DenseNet121 models exhibit reasonable classifier results. However, the MPADL-LCC technique highlights superior performance with a maximum accuracy of 98.87%, precision of 98.14%, sensitivity of 98.87%, and specificity of 98.87%.

These outcomes highlighted the improved performance of the MPADL-LCC approach over other existing systems.

This study presents a novel MPADL-LCC methodology for detecting and classifying leukemia cancer on microscopic images. The presented MPADL-LCC algorithm primarily implements the BF method for the noise removal process. In addition, the MPADL-LCC algorithm employs the Faster SqueezeNet model to develop feature vectors with MPA as a hyperparameter optimizer. The DAE system is exploited for medical image classification at the final stage. The hyperparameter tuning process using MPA helps accomplish enhanced leukemia cancer classification performance. The series of simulations of the MPADL-LCC algorithm is tested on the medical imaging database, and the outcome demonstrates the best results of the MPADL-LCC technique over other recent approaches concerning various measures. In the future, the efficiency of the MPADL-LCC method will be boosted by hybrid DL classification models.

There are several promising avenues to integrate the Marine Predators Algorithm with Deep Learning for Leukemia Cancer Classification on Medical Images in future endeavors. It is crucial to enhance the algorithm’s adaptability to diverse datasets and improve its performance across various leukemia subtypes. Exploring real-time applications in clinical settings and integrating interpretability features into the model can heighten its practical utility. Regular updates are essential to incorporate advancements in deep learning and medical imaging technologies, ensuring the algorithm's relevance in leukemia diagnosis research. Collaboration with medical experts and rigorous validation through extensive clinical trials will solidify its applicability and facilitate seamless integration into healthcare systems.