AML, ALL, and CML classification and diagnosis based on bone marrow cell morphology combined with convolutional neural network: A STARD compliant diagnosis research.

Medicine 2020 AI 10 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
The Problem with Manual Leukemia Diagnosis

Leukemia is a cancer of blood-forming tissues, and its accurate diagnosis depends on examining bone marrow cells under a microscope. Globally, leukemia accounts for about 2.4% of all new cancer cases and 3.2% of cancer deaths annually, making it the 10th most deadly cancer type worldwide.

The three most common leukemia types studied here are AML (Acute Myelogenous Leukemia), ALL (Acute Lymphoblastic Leukemia), and CML (Chronic Myelocytic Leukemia). Each requires different treatment, so accurate classification is critical from the very start.

The current standard method - having a trained hematologist manually examine bone marrow smears under a microscope - is time-consuming, labor-intensive, and subject to human error. Differences in staining quality, lighting, and individual interpretation can all lead to misdiagnosis.

This study proposes using convolutional neural networks (CNNs), a form of artificial intelligence that excels at image recognition, combined with transfer learning, to build a faster, more consistent, and more accurate automated diagnostic system.

TL;DR: Manual leukemia diagnosis from bone marrow images is error-prone, and this study set out to replace it with an AI-based system.
Pages 2-3
Study Design and Patient Data

The researchers collected bone marrow smear images from 104 subjects across three hospitals in China: 18 healthy individuals, 53 AML patients, 23 ALL patients, and 18 CML patients. All samples were taken at initial diagnosis, before any treatment, to avoid confounding the cell appearance.

High-resolution microscopy images (1920x1200 pixels) were captured using a 100x magnification microscope paired with a CCD camera, yielding 1,322 individual cell images in total. The dataset was split 3:1 into a training set (991 images) and a test set (331 images).

Diagnosis for all patients was established by hematologists with at least 15 years of experience, using standard WHO and French-American-British classification criteria, plus additional immunologic and genetic markers. This gave the AI a trustworthy ground truth to learn from.

Three CNN architectures were compared: Inception-V3 (GoogLeNet), ResNet50, and DenseNet121. These are all well-established deep learning frameworks that have won major image recognition competitions and are adapted here for medical imaging.

TL;DR: Images from 104 bone marrow smear samples, classified by expert hematologists, were used to train and test three different neural network architectures.
Pages 3-4
How the Neural Networks Work

A Convolutional Neural Network (CNN) is a type of artificial intelligence designed to process images. Rather than relying on manually defined rules, it learns to detect patterns, textures, and shapes directly from the training images, similar to how a human expert learns over years of practice.

DenseNet121 uses a concept called 'skip connections', where every layer in the network receives input from all preceding layers. This prevents information from being lost as it passes through the network and avoids a common training problem called 'vanishing gradient'.

ResNet50 uses residual learning, which also uses shortcuts between layers to preserve information. Inception-V3 uses multiple filter sizes in parallel within each layer to capture features at different scales simultaneously.

All three networks were configured with the same training settings: a learning rate of 0.001, batch size of 8 images, and 50 training cycles (epochs). The cross-entropy loss function was used to measure how far off each prediction was from the correct answer, guiding the network to improve.

TL;DR: Three proven neural network architectures were trained to recognize bone marrow cell images and classify them into four categories: AML, ALL, CML, or healthy.
Pages 4-5
Transfer Learning - Teaching the AI Faster

Transfer learning is a technique where a neural network trained on one large task is repurposed for a new, related task. Instead of starting from scratch, the model begins with patterns already learned from millions of images, then fine-tunes itself on the smaller medical dataset.

The researchers used models pre-trained on ImageNet, a dataset of over a million natural photographs across 1,000 categories. While bone marrow cells differ from everyday photos, the low-level features learned - edges, textures, color gradients - remain useful.

Two transfer learning strategies were tested. CNN-1 treated the pre-trained model as a fixed feature extractor, only training the final classification layer. CNN-2 used 'fine-tuning', updating all layers with the new medical data, giving the model more flexibility to adapt.

Transfer learning also dramatically reduced training time. The base CNN with no transfer learning took 45 minutes to train; CNN-1 took just 8 minutes; and CNN-2, the most powerful approach, took 20 minutes - less than half the time of the baseline.

TL;DR: By starting with a model pre-trained on millions of natural images, the researchers dramatically cut training time and improved accuracy on the small medical dataset.
Page 5
Image Preprocessing to Improve Quality

Raw bone marrow microscopy images contain many background elements such as red blood cells, platelets, and debris that can confuse the neural network. Sample quality also varied due to differences in staining, storage duration, and lighting conditions.

Two preprocessing steps were applied: the perfect reflection algorithm, which corrects for uneven color and lighting across the image, and a self-adaptive filter, which reduces noise while preserving the structural details of the target cells.

After preprocessing, the best model's accuracy improved from 70.6% to 74.8% - a meaningful but limited gain. The researchers noted that the largest performance jump came from transfer learning, not preprocessing alone.

The preprocessing approach used here is simpler than earlier methods, which required explicitly segmenting each individual cell before analysis. Segmentation is time-consuming and can itself introduce errors in cell shape, so avoiding it streamlined the overall pipeline.

TL;DR: Image preprocessing corrected for staining and lighting variation, providing a modest accuracy boost before the more powerful transfer learning step.
Pages 5-6
Accuracy Results Across All Models

Without transfer learning, the three models achieved test set accuracies of 64.3% (Inception-V3), 66.2% (ResNet50), and 70.6% (DenseNet121) on raw images. After preprocessing, these became 60.6%, 69.3%, and 74.8% respectively - DenseNet121 consistently led.

Adding transfer learning to DenseNet121 produced striking results. The CNN-2 fine-tuning model achieved 95.3% overall accuracy, a gain of 20.5 percentage points over the preprocessed-only DenseNet121. CNN-1 (feature extraction only) achieved 84.9%.

Per-category accuracy for the best model (CNN-2) was: AML - 90%, ALL - 97%, CML - 95%, Healthy - 99%. AML had the lowest accuracy because its immature granulocytes visually resemble the immature lymphocytes seen in ALL, making them harder to distinguish.

The confusion matrix revealed that 8% of AML images were incorrectly labeled as ALL, and 2% were labeled as healthy. This reflects a genuine biological challenge: AML and ALL both feature large numbers of immature, rapidly dividing cells that share overlapping visual characteristics.

TL;DR: The final AI model achieved 95.3% accuracy in classifying leukemia type from bone marrow images, far surpassing results without transfer learning.
Pages 6-7
Why AML is Hardest to Classify

AML (Acute Myelogenous Leukemia) shows the most misclassification, primarily being confused with ALL. Both cancers feature a large number of immature, rapidly dividing blast cells. In AML, these are immature granulocytes and monocytes; in ALL, they are immature lymphocytes.

The key differences between AML and ALL blast cells are subtle - differences in nuclear texture, degree of staining, and cytoplasm-to-nucleus ratio. These fine distinctions are what both expert pathologists and the AI must learn to detect reliably.

CML and healthy bone marrow are easier to distinguish because they primarily contain mature cells, which have more distinctive, stable morphology than the rapidly changing immature blasts seen in acute leukemias.

The researchers note that even human experts with decades of experience can misclassify leukemia subtypes based on morphology alone, which is why clinical diagnosis also includes genetic and immunological testing. The AI's 90% AML accuracy is therefore a strong result for a vision-only system.

TL;DR: AML is most often confused with ALL because both diseases share similar populations of immature blast cells that differ only in subtle visual details.
Pages 6-7
Bone Marrow vs. Peripheral Blood Images

Most prior studies in this area used peripheral blood smear images, which are easier to collect but less informative. This study specifically used bone marrow smear images, which more directly reflect the underlying biology of the disease.

Bone marrow smears show the full process of blood cell synthesis, maturation, and release within the marrow itself. This makes them the gold standard for leukemia diagnosis, as they capture disease-specific cell populations that may not appear in peripheral blood at all.

Earlier AI studies were also typically limited to one leukemia type at a time. This study's system simultaneously distinguishes between three different types of leukemia and healthy controls in a single model, making it significantly more practical for clinical use.

The researchers also note that the same CNN method could potentially be applied to other bone marrow examination types, including bone marrow imprints and bone marrow sections, broadening its potential clinical utility beyond the specific smear format used here.

TL;DR: Unlike most prior work, this study used bone marrow rather than peripheral blood images, and classified three leukemia types simultaneously rather than one at a time.
Page 7
Clinical Potential and Limitations

The most immediate clinical benefit of this system would be in settings where expert hematologists are scarce. The AI model produces consistent, objective results in under 20 minutes of training time and can classify images far faster than manual review.

The system does not require cell segmentation, which means it can work with standard microscopy images without specialized preprocessing equipment. This lowers the barrier to clinical adoption and could make high-quality leukemia screening more accessible.

A significant limitation is the small dataset size. With only 104 patients, the model may not generalize perfectly to the full diversity of leukemia presentations seen in a real clinic. Larger, more diverse datasets would be needed before clinical deployment.

The study did not include Chronic Lymphocytic Leukemia (CLL), which has low incidence in China but is common in Western countries. Future work is planned to expand the model to cover more leukemia types and to validate against the full WHO classification system, including genetic markers.

TL;DR: The AI model shows strong clinical promise but needs larger, more diverse training datasets and broader leukemia type coverage before real-world deployment.
Page 7
Key Takeaways

This study demonstrates that a DenseNet121 CNN combined with transfer learning fine-tuning can classify three major leukemia types from bone marrow cell images with 95.3% accuracy, a level competitive with human expert performance for a vision-based system.

Transfer learning was the single most impactful technical choice, boosting accuracy by over 20 percentage points and reducing training time by more than half. This makes the approach practical even when labeled medical data is limited.

The model provides rapid, objective, and reproducible diagnoses that avoid the human subjectivity inherent in manual smear reading. This could help standardize leukemia diagnosis across hospitals with varying levels of hematology expertise.

Looking ahead, the authors plan to expand this framework to cover more leukemia subtypes and to incorporate multi-modal data, combining image analysis with genetic and immunological markers for even higher diagnostic precision.

TL;DR: AI-powered bone marrow image analysis achieved 95.3% leukemia classification accuracy and could one day help standardize and speed up diagnosis worldwide.
Citation: Open Access, 2020. Available at: PMC7647529.