Leukemia is a cancer of the blood and bone marrow characterized by uncontrolled production of abnormal white blood cells (WBCs). These malignant cells crowd out healthy blood cells, weakening the immune system and impairing the body's ability to fight infection, carry oxygen, and clot blood.
Leukemia is classified along two axes: how fast it progresses and which type of white blood cell is affected. Acute leukemia progresses rapidly - the abnormal cells are so immature they cannot perform normal functions. Chronic leukemia progresses slowly and the cells retain some normal function, making it harder to detect because the cells can appear similar to healthy ones.
Each type is further divided by the cell lineage involved. Lymphoid leukemias arise from lymphocytes (immune cells), while myeloid leukemias arise from other blood cell precursors. This gives the four main subtypes: Acute Lymphocytic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myeloid Leukemia (CML).
Correctly identifying the specific subtype is critical because each responds differently to treatment. The wrong diagnosis can lead to ineffective therapy and worse patient outcomes. This makes automated, accurate subtype identification an important clinical goal.
The primary diagnostic method for leukemia involves examining a blood smear under a microscope - a thin layer of blood spread on a glass slide and stained so that individual cells are visible. Trained hematologists look for abnormal cell morphology: unusual sizes, shapes, and nuclear structures that indicate malignancy.
This process is time-consuming, requires highly trained specialists, and is subject to human fatigue and variability. When a physician needs to assess hundreds of cells per patient, and results must be available quickly, the manual approach creates a bottleneck. Automated systems could dramatically speed up diagnosis and reduce error rates.
Previous machine learning approaches achieved high accuracy (up to 99%) for simple binary classification - distinguishing one leukemia subtype from healthy cells. However, distinguishing all four leukemia subtypes from each other and from healthy cells simultaneously is a much harder, multi-class problem that had not yet been solved. This study addresses that gap directly.
The researchers assembled a dataset by combining two publicly available sources. The ALL-IDB dataset contains annotated microscopic blood images of ALL leukemia and healthy cells, where expert oncologists labeled each cell's classification - making it highly reliable. However, it only contains ALL and healthy samples.
The ASH Image Bank (from the American Society of Hematology) contains a broader collection including all four leukemia subtypes. The combined original dataset contained 179-185 images per subtype - far too few for deep learning, which typically requires thousands of examples to learn meaningful patterns reliably.
The original totals were 354 images from ALL-IDB and 549 from ASH Image Bank. While these numbers are reasonable for traditional machine learning, convolutional neural networks (CNNs) need much larger training sets to avoid memorizing the training data (overfitting) rather than learning generalizable features. Data augmentation was essential to make deep learning viable.
Data augmentation is the practice of creating modified versions of existing images to expand the training set. The key insight is that a rotated or flipped blood cell image still shows the same type of leukemia - the transformation doesn't change the underlying biology. By training on many variations, the model learns to recognize the fundamental patterns rather than memorizing specific image artifacts.
The researchers applied seven distinct image transformations: rotation (40 degrees), height shift (40%), width shift (40%), zoom (30%), horizontal flip, vertical flip, and shearing (20 degrees). Each original image was transformed using randomly selected combinations of these operations.
This augmentation expanded the dataset by a factor of 8. The ALL-IDB dataset grew from 354 to 2,478 images, and the ASH Image Bank dataset grew from 549 to 3,843 images. This provided the CNN with enough diverse examples to learn robust, generalizable features for each leukemia subtype without overfitting to any single image.
A convolutional neural network (CNN) is a type of deep learning model specifically designed for image recognition. Rather than treating each pixel independently, CNNs apply learnable filters that slide across the image, detecting patterns like edges, textures, and shapes at multiple spatial scales - mimicking how the visual cortex processes images.
The architecture used in this study had two convolutional layers, each with 32 feature maps using 3x3 filters, followed by 2x2 max-pooling layers that downsample the feature maps while retaining the most important information. This design was intentionally kept compact to reduce overfitting on the limited dataset.
After the convolutional layers, a flattening layer converts the 2D feature maps into a 1D vector, which feeds into a fully connected neural network with 128 nodes. Dropout (randomly disabling 10% of neurons during training) and batch normalization were applied to reduce overfitting. The output layer has 5 nodes - one for each leukemia subtype and one for healthy - controlled by a SoftMax function that outputs a probability for each class.
Two optimizers were tested: SGD (stochastic gradient descent) and ADAM. The model was trained for 25 epochs (full passes through the training data) with a batch size of 32, selected after experiments showed longer training did not improve performance.
The best binary classification result was achieved with data augmentation: 88.25% validation accuracy for distinguishing ALL from healthy cells (Experiment 2). Without augmentation, the same task achieved only 81.16%, confirming that augmentation is crucial for CNN performance on small medical image datasets.
For the harder multi-class problem (distinguishing all four leukemia subtypes and healthy cells), the best result was 81.74% accuracy using the SGD optimizer with 25 epochs. Interestingly, training for 100 epochs rather than 25 caused performance to drop to 66.41%, indicating that longer training leads to overfitting despite the augmented dataset size. The ADAM optimizer performed even worse at 63.40%.
The CNN significantly outperformed all traditional machine learning approaches on both tasks. For binary classification, naive Bayes reached 69.69%, decision tree 62.94%, k-nearest neighbor 58.57%, and SVM only 50.09% - barely better than chance. The gap was even larger for multi-class classification, where SVM collapsed to just 20.84% while the CNN achieved 81.74%.
Performance varied substantially across the 5 cross-validation folds, with individual fold accuracy ranging from 65% to 98%. This variability highlights the sensitivity of the results to which specific images end up in the training versus test sets, and underscores the importance of cross-validation to get reliable average performance estimates.
To the authors' knowledge, this is the first study to attempt automated classification of all four leukemia subtypes simultaneously from microscopic blood images. Previous studies either focused on binary classification (one subtype vs. healthy) or on sub-classifying subtypes within a single leukemia type (e.g., different stages of ALL).
The multi-class problem is fundamentally harder because different leukemia subtypes can share visual features. For example, the distinction between ALL and CLL involves subtle differences in lymphocyte morphology that even experienced hematologists may find challenging. The fact that the CNN achieved over 80% accuracy across all five classes is a meaningful proof of concept.
Compared to the published literature, the study performed slightly below the top binary classifiers (Shafique et al. achieved 99.5%, Thanh et al. achieved 96.6%), but those studies did not use cross-validation, making their reported accuracy estimates potentially optimistic. When evaluated rigorously with cross-validation, the CNN in this study achieved competitive performance.
This study demonstrated that a CNN trained with data augmentation can reliably distinguish all four major leukemia subtypes from each other and from healthy blood cells, achieving 88% accuracy for binary and 81% accuracy for multi-class classification using rigorous cross-validation.
The results confirm that CNNs substantially outperform traditional machine learning algorithms (SVM, Naive Bayes, Decision Trees, k-NN) for this image classification task. The gap is particularly dramatic for multi-class leukemia subtyping, where traditional methods struggle but CNNs maintain reasonable performance.
Future work will explore hybrid architectures combining CNNs with recurrent neural networks (RNNs), which could capture sequential patterns in blood cells. The researchers also plan to expand the dataset by collecting new samples and applying additional augmentation techniques, which would likely improve both binary and multi-class accuracy toward clinically useful thresholds.