What is leukemia? Leukemia is a cancer that affects the white blood cells (WBC), the body's infection-fighting cells. When leukemia develops, the bone marrow begins producing large numbers of immature, abnormal white blood cells that crowd out healthy cells and weaken the immune system.
Types of leukemia: Leukemia is broadly divided into four subtypes based on how fast it progresses and which type of white blood cell is affected: Acute Lymphocytic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myeloid Leukemia (CML). Acute forms grow and spread quickly, while chronic forms progress more slowly.
Diagnosis challenge: Identifying leukemia - and then pinpointing its specific subtype - is critical because treatment plans differ significantly between subtypes. However, the traditional method of manually examining blood smear images under a microscope is time-consuming, requires a specialist, and is prone to human error.
Why early detection matters: Early and accurate identification of leukemia type directly impacts the chance of successful treatment. Delays or misclassification can lead to inappropriate therapy, worsening outcomes for patients.
The Internet of Medical Things (IoMT): IoMT refers to a network of smart medical devices, sensors, and applications connected to healthcare IT systems via the internet. These devices collect patient data and transmit it to cloud platforms for analysis, enabling remote monitoring and diagnosis.
The proposed system: This study introduces an IoMT-based framework where an IoT-enabled microscope automatically captures blood smear images and uploads them to a secure cloud platform called the "leukemia cloud." Deep learning models then analyze these images and return a diagnosis - all without requiring an on-site expert.
Real-time coordination: The system links patients, devices, and healthcare providers in real time. After the AI classifies the leukemia subtype, results are displayed on the clinician's computer, allowing the doctor to recommend treatment remotely. This approach could significantly reduce delays in rural or underserved areas.
Pandemic relevance: The framework was also designed with crisis situations in mind. During the COVID-19 pandemic, many patients with chronic diseases were reluctant or unable to travel to hospitals. This remote IoMT system would allow safe leukemia diagnosis and care from home, maintaining quarantine while ensuring patients still receive attention.
What is deep learning? Deep learning is a branch of artificial intelligence that uses layered networks of mathematical operations to automatically learn patterns from raw data. Unlike traditional methods that require experts to manually define which image features to analyze, deep learning identifies these patterns on its own.
ResNet-34 explained: ResNet-34 is a 34-layer deep neural network that solves a common problem in very deep networks called the "vanishing gradient" - a phenomenon where training becomes ineffective as networks grow deeper. ResNet addresses this by creating shortcut connections that allow information to flow directly between distant layers, keeping training efficient even at depth.
DenseNet-121 explained: DenseNet-121 is a 121-layer network with a unique architecture where every layer is connected to every subsequent layer in the network. This dense connectivity allows each layer to receive and use features learned by all previous layers, making the model very efficient at learning from smaller datasets and less prone to overfitting.
Transfer learning: Both models were applied using transfer learning, meaning they were initially trained on a large general image dataset (ImageNet) and then fine-tuned for the specific task of leukemia classification. This approach enables high accuracy even when the available medical image dataset is relatively small.
Data sources: Two publicly available datasets were used: the ALL-IDB (Acute Lymphoblastic Leukemia Image Database), which contains annotated blood cell images classified by expert oncologists, and the ASH Image Bank from the American Society of Hematology, which includes images covering all four leukemia subtypes.
Dataset composition: Before augmentation, the combined dataset contained 181 ALL images, 55 AML images, 38 CLL images, 57 CML images, and 187 healthy cell images - a relatively small collection that could limit deep learning performance.
Data augmentation: To expand the training data and prevent overfitting, the researchers applied image transformation techniques including rotation, horizontal flipping, height and width shifting, zooming, and shearing. These techniques create varied versions of each original image, giving the AI model a broader range of examples to learn from.
Augmented dataset size: After augmentation, the ASH image bank dataset grew to 3,277 images and the ALL-IDB dataset reached 2,359 images. This expansion allowed the deep learning models to train more robustly and generalize better to new, unseen blood cell images.
Evaluation metrics: The models were assessed using four standard metrics: accuracy (overall correct predictions), precision (how often a positive prediction is correct), recall (how well the model detects actual cases), and F1 score (a combined measure of precision and recall). A confusion matrix was also produced to visualize classification errors for each leukemia subtype.
ResNet-34 results: ResNet-34 achieved 100% accuracy for ALL detection and healthy classification. For AML, CLL, and CML, accuracy reached 99.65%, 99.73%, and 99.73% respectively - all with precision, recall, and F1 scores above 0.99. The overall average accuracy across all subtypes was 99.56%.
DenseNet-121 results: DenseNet-121 outperformed ResNet-34, achieving 100% accuracy for ALL, CML, and healthy cases. AML and CLL accuracy reached 99.91% each, with precision, recall, and F1 scores all at or near 1.0. The overall average accuracy was 99.91%.
Why DenseNet-121 wins: The dense connection architecture of DenseNet-121 allowed it to extract richer features from the blood cell images with fewer parameters, resulting in lower training and validation loss and better generalization across all leukemia subtypes compared to ResNet-34.
Prior limitations: Most earlier computer-aided diagnosis studies for leukemia focused only on one or two subtypes - typically distinguishing healthy cells from ALL, or comparing ALL and AML. None had successfully classified all four leukemia subtypes (ALL, AML, CLL, CML) simultaneously in a single model.
Comparison with previous methods: Earlier approaches using traditional machine learning such as support vector machines (SVM), K-nearest neighbor (KNN), and basic convolutional neural networks achieved lower accuracy. A standard CNN achieved only 81.74% accuracy for subtype classification, and even a genetic algorithm combined with SVM reached only 99.50% on a more limited classification task.
Improvement demonstrated: DenseNet-121 achieved 99.91% average accuracy across all five categories (four leukemia subtypes plus healthy), and ResNet-34 reached 99.56%. Critically, both models accomplished the task of classifying all four subtypes simultaneously - something no previous study in this comparison had done.
No hand-crafted features needed: Unlike older machine learning methods that required researchers to manually design and extract specific image features (such as color histograms or texture measures), the proposed deep learning models learn relevant features automatically from raw image data, saving significant time and expert effort.
Reducing expert dependency: Diagnosing leukemia currently requires trained hematologists to spend significant time examining blood smear slides under a microscope. The proposed IoMT framework automates this process, potentially enabling diagnosis in settings where specialist access is limited or unavailable.
Remote healthcare access: By connecting an IoT-enabled microscope to a cloud AI system, clinicians in remote hospitals or community clinics could receive expert-level leukemia classification results instantly, without needing a specialist physically present. This is particularly valuable in developing countries or rural areas.
Pandemic and crisis scenarios: The COVID-19 pandemic highlighted how fragile in-person healthcare delivery can be. The IoMT framework enables leukemia patients to receive diagnosis and monitoring from home, avoiding unnecessary hospital visits and reducing exposure risks for both patients and healthcare workers.
Faster treatment decisions: Because the system produces results quickly and with high accuracy, doctors can begin appropriate treatment plans sooner. In acute leukemia cases where rapid progression is common, even days saved in diagnosis time can meaningfully improve a patient's prognosis.
Larger and more diverse datasets: The authors suggest that future work should incorporate larger blood cell image collections with greater variety in patient demographics, imaging equipment, and staining methods. A more diverse dataset would improve the model's ability to work across different clinical settings.
Subcategory classification: Each major leukemia type has further subtypes. For example, ALL is divided into L1, L2, and L3 subtypes, and AML has eight further classifications. Future versions of the framework could be extended to diagnose these finer-grained categories, providing even more specific information to guide treatment choices.
Other blood abnormalities: Because the deep learning models analyze blood cell images directly, the same framework could be adapted to identify other abnormalities in blood beyond leukemia, such as anemia, infections, or other hematological disorders - broadening its clinical utility.
Integration with treatment systems: Linking the diagnostic AI to electronic health records and treatment planning systems would create a more complete digital healthcare pathway, from initial diagnosis through treatment monitoring, potentially improving long-term outcomes for leukemia patients.