White blood cell (WBC) leukemia is caused by the uncontrolled proliferation of abnormal leukocytes in the bone marrow, which then circulate in the blood. Early and accurate identification of malignant WBCs is critical for diagnosis and treatment planning. Traditionally, this is done by trained hematologists examining blood smear images under a microscope - a slow, labor-intensive process prone to human error and fatigue.
WBCs exist in multiple types including granulocytes (neutrophils, eosinophils, basophils) and agranulocytes (lymphocytes, monocytes). In acute lymphoblastic leukemia (ALL), malignant lymphocytes called lymphoblasts appear under the microscope with characteristic features: irregular nuclear envelopes, small cavities in the cytoplasm called vacuoles, and small round particles called nucleoli within the nucleus. However, the visual distinction between benign and malignant cells can be subtle and difficult.
Automated image classification using deep learning - particularly convolutional neural networks (CNNs) - represents the current state of the art for this type of medical imaging task. CNNs learn to extract meaningful patterns from large image datasets and can match or exceed human performance on classification tasks. However, CNNs trained from scratch require massive datasets and significant computational resources.
A practical approach for medical imaging, where labeled data is typically scarce, is transfer learning: using a CNN pre-trained on millions of general images (such as the ImageNet dataset) and adapting it to the specific medical task. The network already knows how to detect general visual features like edges, textures, and shapes, and these features can be repurposed for detecting cancer cell morphology.
The researchers used VGGNet, one of the most powerful CNN architectures, pre-trained on the ImageNet dataset of over 1.2 million images across 1,000 categories. VGGNet consists of 16 convolutional layers and 3 fully connected layers. By removing the final classification layer, the network becomes a feature extractor: any image passed through it is converted into a numerical feature vector describing its visual content.
For each WBC image of size 224 x 224 pixels, VGGNet generates a feature vector of 25,088 numbers, representing a rich description of the image's visual patterns. However, this enormous feature space is highly redundant - many features capture similar or correlated information, and many contain noise rather than useful signal. Classifying images directly from 25,088 features is inefficient and can even reduce accuracy through overfitting.
The core challenge this paper addresses is feature selection: identifying which of the 25,088 VGGNet features are actually informative for distinguishing malignant from benign WBCs, and discarding the rest. Effective feature selection reduces computational cost, speeds up training and deployment, and often improves classification accuracy by removing confounding noise.
Two public benchmark datasets were used for testing: ALL-IDB2, containing 260 images (50% benign, 50% malignant), and C-NMC, a larger dataset of 10,661 images from the B-ALL classification challenge at IEEE ISBI-2019 (7,272 from ALL patients, 3,389 from healthy subjects). The 80/20 training/test split with 5-fold cross-validation ensured rigorous evaluation.
To solve the feature selection problem, the researchers developed the Statistically Enhanced Salp Swarm Algorithm (SESSA). The basic Salp Swarm Algorithm (SSA) is a bio-inspired optimization method that mimics the foraging behavior of salps - small marine invertebrates that move in coordinated chains to navigate the ocean in search of food.
In SSA, the population of candidate feature subsets is divided into leaders (front salps that orient toward the best-known solution) and followers (remaining salps that trail behind, maintaining population diversity). Over many iterations, the swarm converges on the feature subset that maximizes classification accuracy while minimizing the number of features selected.
The enhanced SESSA version adds three statistical pre-processing steps before the swarm optimization begins: (1) chi-square testing to remove correlated features, (2) recursive feature elimination to remove features with low predictive power, and (3) tree-based feature importance ranking. These steps dramatically narrow the search space before the swarm even begins, making the optimization faster and more reliable.
The combination of statistical filtering and swarm-based search exploits the complementary strengths of both approaches. Statistical methods quickly eliminate obviously useless features, while the swarm algorithm explores the remaining feature space intelligently to find the best subset - achieving better results than either approach alone.
On the ALL-IDB2 dataset, SESSA selected only 1,087 features out of the 25,088 provided by VGGNet - just 4% of the original feature set. Despite using 96% fewer features, the proposed approach achieved 96.1% accuracy, compared to 94.2% for VGGNet alone. Sensitivity reached 99.6% and specificity 92.9%.
On the larger C-NMC dataset, SESSA again selected approximately 4% of features and achieved 83.3% accuracy versus 80.9% for VGGNet with all features. While the improvement is smaller on this more challenging dataset, the combination of higher accuracy with dramatically fewer features represents a meaningful advance in efficiency.
Comparing SESSA against four other optimization algorithms - statistically enhanced versions of multi-verse optimization, grey wolf optimization, particle swarm optimization, and a genetic algorithm - SESSA achieved the lowest classification error and the highest accuracy across most metrics on both datasets. This confirms that the salp swarm approach is particularly well-suited to this type of feature selection problem.
Even the worst-performing fold in the cross-validation rarely dropped below 80% accuracy for most classifiers, demonstrating that the performance is robust and not dependent on a particularly favorable train-test split. Six different classifiers (linear SVM, KNN, decision trees, naive Bayes, AdaBoost, and multi-layer perceptron) were used to evaluate the selected features, confirming the selected features are genuinely informative rather than classifier-specific.
The clinical motivation for this work is straightforward: if leukemia diagnosis can be automated from microscopy images with high accuracy, it becomes faster, cheaper, and less dependent on specialized expertise. This matters especially in resource-limited healthcare settings where experienced hematologists are scarce.
By reducing the feature space by 96%, the proposed approach also reduces computational costs proportionally. Models with fewer features train faster, require less memory, and can be deployed on less powerful hardware - potentially enabling leukemia screening in clinics that cannot afford high-end computing infrastructure.
The combination of CNN feature extraction (which handles the complexity of learning from raw images) with intelligent feature selection (which removes redundancy and noise) offers a broadly applicable framework for medical image classification beyond leukemia. Any classification task where a powerful pre-trained CNN extracts too many features could benefit from this approach.
The authors note that the proposed approach is intended as a methodological contribution rather than a ready-to-deploy clinical diagnostic tool. Further validation on larger, more diverse patient cohorts and integration with clinical workflows would be required before such a system could be used in routine practice.
The SESSA-VGGNet hybrid approach establishes a new performance benchmark on both the ALL-IDB2 and C-NMC leukemia classification datasets, outperforming previous methods including those based on larger CNN architectures and earlier feature selection methods. This positions it as a competitive approach for the WBC classification task.
The paper demonstrates a key insight: for specialized classification tasks like medical image analysis, more features are not always better. Careful selection of the most relevant features - even if they represent a tiny fraction of all available features - produces better, faster, and more reliable models than using every available feature.
Future work could extend this approach in several directions: applying it to other leukemia subtypes, combining image features with clinical metadata, or using it as a pre-processing step before feeding features into more sophisticated deep learning architectures. The public availability of the code and the reproducibility of the approach support further development by the research community.
As artificial intelligence tools for medical imaging continue to mature, efficient approaches like SESSA-VGGNet that balance performance with computational practicality will be essential for real-world clinical deployment - particularly in global health contexts where technological resources are limited but cancer burden is high.