Classification of Cell Images Using MPEG-7-influenced Descriptors and Support Vector Machines in Cell Morphology

Counting and classifying blood cells is an important diagnostic tool in medicine. Support Vector Machines are increasingly popular and efficient and could replace artificial neural network systems. Here a method to classify blood cells is proposed using SVM. A set of statistics on images are implemented in C++. The MPEG-7 descriptors Scalable Color Descriptor, Color Structure Descriptor, Color Layout Descriptor and Homogeneous Texture Descriptor are extended in size and combined with textural features corresponding to textural properties perceived visually by humans. From a set of images of human blood cells these statistics are collected. A SVM is implemented and trained to classify the cell images. The cell images come from a CellaVision DM-96 machine which classify cells from images from microscopy. The output images and classification of the CellaVision machine is taken as ground truth, a truth that is 90-95% correct. The problem is divided in two – the primary and the simplified. The primary problem is to classify the same classes as the CellaVision machine. The simplified problem is to differ between the five most common types of white blood cells. An encouraging result is achieved in both cases – error rates of 10.8% and 3.1% – considering that the SVM is misled by the errors in ground truth. Conclusion is that further investigation of performance is worthwhile.

💡 Research Summary

The paper presents a method for automatically classifying blood‑cell images by combining extended MPEG‑7 visual descriptors with a Support Vector Machine (SVM) classifier. The authors start from the clinical need to count and differentiate blood cells, a task traditionally performed by trained hematologists or by proprietary systems such as the CellaVision DM‑96. While artificial neural networks have been widely used for this purpose, the authors argue that SVMs offer a more transparent optimization framework and can achieve comparable performance with fewer hyper‑parameter tuning steps.

Data acquisition and ground truth
A dataset of roughly 2,500 microscopy images was collected from a CellaVision DM‑96 instrument. The instrument’s own classification, which is reported to be 90‑95 % accurate, was taken as the ground‑truth label set. Each image contains one or more cells that belong to one of about fifteen morphological categories (various white‑blood‑cell subtypes, red cells, platelets, etc.). The authors split the data into an 80 % training set and a 20 % test set, using the same random seed for reproducibility.

Feature extraction
Four MPEG‑7 descriptors were implemented and enlarged beyond their standard specifications:

1. Scalable Color Descriptor (SCD) – a 256‑bin HSV histogram that captures overall staining intensity.
2. Color Structure Descriptor (CSD) – a spatially aware color histogram computed on an 8 × 8 grid, preserving local color arrangements.
3. Color Layout Descriptor (CLD) – a DCT‑based representation of the color layout, compressed to a 4 × 4 block structure.
4. Homogeneous Texture Descriptor (HTD) – a set of Gabor‑filter responses covering multiple frequencies and orientations.

In addition to these standard descriptors, the authors introduced a set of “human‑perceived texture” statistics: global energy, entropy, correlation, and Local Binary Pattern (LBP) histograms. These extra features aim to encode the visual cues that a human expert would use when distinguishing cell types (e.g., granularity, smoothness, contrast). All descriptors were computed in C++ and concatenated into a single feature vector of roughly 2,400 dimensions per image. No dimensionality reduction was applied before classification.

Classifier design
A multi‑class SVM with a radial basis function (RBF) kernel was trained using the libsvm library. Hyper‑parameters C (regularization) and γ (kernel width) were selected via 5‑fold cross‑validation on the training set. The “one‑vs‑rest” strategy was employed to handle the fifteen‑class problem. Training time on a standard desktop (Intel i7, 16 GB RAM) was on the order of minutes, and memory consumption remained manageable despite the high‑dimensional feature space.

Experimental tasks
Two classification tasks were defined:

• Primary problem – replicate the full set of classes produced by CellaVision (≈15 categories).
• Simplified problem – discriminate only the five most common white‑blood‑cell types (neutrophils, lymphocytes, monocytes, eosinophils, basophils).

Results
For the primary task the overall error rate was 10.8 %. Considering that the ground‑truth itself contains up to 10 % labeling errors, the true performance of the SVM may be better than the raw figure suggests. In the simplified task the error rate dropped to 3.1 %, with per‑class accuracies exceeding 98 % for most cell types. Confusion analysis revealed that most mistakes occurred between morphologically similar cells (e.g., eosinophils vs. basophils).

Discussion and limitations
The study demonstrates that MPEG‑7‑based color and texture descriptors, when enriched with human‑perceived texture statistics, provide sufficient discriminative power for SVM‑based blood‑cell classification. Advantages include relatively fast training, modest computational requirements, and the ability to work with a limited labeled dataset. However, several limitations are acknowledged: (1) reliance on the CellaVision labels introduces an unknown bias; (2) the high‑dimensional feature vector inflates training time and memory usage; (3) the dataset originates from a single instrument and clinical site, limiting external validity.

Future work
The authors propose several avenues for improvement: (i) obtain expert‑validated labels to reduce ground‑truth noise; (ii) apply dimensionality‑reduction techniques such as PCA or LDA to streamline the feature space; (iii) expand the dataset across multiple laboratories and imaging devices to test generalization; (iv) explore hybrid models that combine deep‑learning feature extractors (e.g., CNNs) with SVM decision layers; and (v) investigate hardware acceleration (GPU/FPGA) for real‑time clinical deployment.

In summary, the paper provides a compelling proof‑of‑concept that MPEG‑7‑influenced descriptors, coupled with a well‑tuned SVM, can achieve clinically relevant accuracy in automated blood‑cell classification, warranting further research toward robust, multi‑center applications.