Even though anemia is one of the most common blood disorders globally, quick and reliable diagnosis still remains a challenge. Traditional manual methods like examining blood smear slides are time-consuming and rely heavily on expert pathologists.
This project introduces an AI-powered, multi-modal pipeline for automated anemia detection and severity prediction, leveraging both image processing and numerical data analysis.
We developed two parallel workflows:
- 🩸 Blood smear images (for detecting anemia)
- 📊 CBC numerical data (for predicting anemia severity)
This hybrid approach helps overcome the limitations of relying on just one type of data — bringing speed, precision, and automation into clinical diagnostics.
“Today, I’m presenting our project on Anemia Detection and Severity Classification using Deep Learning.
Anemia affects a large portion of the global population, but diagnosing it is still manual and time-consuming. Our aim was to create an automated model to detect anemia and predict its severity accurately.
We used a multi-modal approach, because we’re working with two types of data:
- Blood smear images
- CBC numerical values (like hemoglobin levels)
For the image side: We used segmented RGB blood smear images, where we removed the noise and isolated only the red blood cells (RBCs). This helped the model focus on what truly matters.
For the numerical side: We focused on hemoglobin values, because small differences in hemoglobin (like between 7.5 and 8 g/dL) are very clinically important. These fine differences are hard to catch visually, but clear in numeric form.
So:
- 🖼️ Images helped us detect anemia
- 📈 Numerical values helped us predict severity
This multi-modal fusion made our model more accurate and realistic for medical use.”
- Format:
.jpgor.png - Channels: RGB Segmented
- Preprocessed to isolate RBCs and remove background, WBCs, platelets
- Classes:
AnemicHealthy
- Columns: Hemoglobin, MCV, MCH, MCHC, RBC Count
- Labels:
MildModerateSevere
💡 Severity label derived from hemoglobin thresholds based on WHO standards.
- Input: 224 × 224 RGB image
- Architecture: CNN or pretrained model (e.g., ResNet)
- Layers: Conv → ReLU → MaxPool → Dropout → Dense → Sigmoid
- Output: Binary classification (Anemic / Healthy)
- Model: Random Forest / SVM / Neural Network
- Input: Hemoglobin and other blood parameters
- Output: Severity level – Mild / Moderate / Severe
- ✅ RGB segmentation to isolate red blood cells
- 📊 Feature-driven severity classification
- 🔀 Data augmentation: rotate, flip, scale (images)
- 📈 K-Fold cross-validation for better generalization
- 📦 Ready-to-run in Google Colab / Jupyter Notebook
| Task | Accuracy |
|---|---|
| Anemia Detection | 92.3% |
| Severity Prediction | 88.7% |
- ROC-AUC Curve plotted
- Confusion Matrix visualized
- Feature importance explained using SHAP (optional extension)
| Tool / Library | Purpose |
|---|---|
| Python | Core programming language |
| TensorFlow / Keras | CNN modeling for image analysis |
| Scikit-learn | Severity prediction model |
| OpenCV | Image preprocessing |
| Pandas / NumPy | Data manipulation |
| Matplotlib / Seaborn | Visualization |
git clone https://github.com/keerthana777z/Anemia-detection-using-Blood-smear-images-.git cd Anemia-detection-using-Blood-smear-images-
pip install -r requirements.txt
jupyter notebook
🌱 Future Enhancements 🧠 Add Grad-CAM for interpretability of image predictions
🌐 Deploy as a Streamlit web app for hospitals/clinics
☁️ Integrate cloud storage for patient record access
🔬 Multi-class classification for anemia types (e.g., iron-deficiency, sickle cell)
👩💻 Author AR Keerthana
📄 License This project is licensed under the MIT License – free to use, improve, and distribute.
“Let’s build healthcare where AI doesn’t just detect, it prevents.” 💡