Leukemia_Detector_And_Predictor_Using_ResNet_Main

Automated Blood Cell Classification Using Deep Learning (ResNet–50)

A lightweight, end-to-end system for detecting and classifying leukemic cells from microscopic blood smear images. This project includes:

-Deep learning model (ResNet-50 backbone)

-Automated cell segmentation using OpenCV

-Flask backend (model inference + preprocessing + segmentation)

-Streamlit frontend (user interface + visualization)

-Logging of predictions and annotated outputs

1. Project Overview

Early detection of leukemia can significantly improve treatment outcomes. This system classifies blood cells into:

-Leukemic

-Normal

The workflow includes:

-Image Upload (JPG/PNG/BMP)

-Automated Nucleus Segmentation

-Bounding-Box Extraction

-ResNet-based Classification

-Probability Scores + Annotated Image Output

-Prediction Logging for Auditability

2. How the Model Performs Classification

The model doesn’t look at the entire image at once. Instead, it:

--Segments the nucleus or dominant WBC region --Using thresholding + contour detection.

--Extracts the region of interest (ROI) --This focuses the model on the actual cell, not the background.

--Resizes and normalizes the ROI (224×224, scaled between 0–1)

--Uses a fine-tuned ResNet-50 to classify the cell.

What ResNet Learns in Leukemia vs Normal Cells

The model focuses on:

-Feature Leukemic Cells Normal Cells -Nucleus Shape Often irregular, larger, more variable More uniform and rounded -Chromatin Texture Dense, clumped Smooth and consistent -Color/Staining Pattern Darker nucleus, uneven staining Evenly stained -Cell Boundary May appear distorted Clear, defined boundary

ResNet automatically extracts multi-level features from these patterns across 50 layers.

3. Performance Metrics

Classification Report (Test Set)

Class Precision Recall F1-Score Support

Leukemia 0.83 0.90 0.87 1092

Normal 0.74 0.61 0.67 509

Accuracy 0.81 1601

4. Interpreting These Results

The model is highly sensitive to leukemia cells (recall 0.90).

Meaning: If a cell is leukemic, there is a 90% chance the model will catch it.

Normal cells have lower recall (0.61), meaning:

Some normal cells are misclassified as leukemic.

This is common when:

The dataset is imbalanced (more leukemia cells than normal)

Normal cell features vary mildly across labs and staining methods

Still, 81% overall accuracy indicates the model generalizes reasonably well.

5. Important Note on Dataset & Hardware Limitations

This project was built under:

Dataset Limitations.

Limited number of normal cell samples.

Some samples had inconsistent staining / imaging quality.

No professionally annotated bounding boxes.

All segmentation was hand-designed (not a learned segmentation model)

These factors naturally limit the maximum achievable accuracy.

Device Limitations:

The model was trained on limited hardware, restricting:

Total training time.

Batch size.

Ability to use more complex architectures.

Ability to perform extensive hyperparameter tuning.

Ability to train with large augmentations.

Future Improvements

-With:

--Larger balanced datasets

--Better quality microscopic images

--Modern architectures (EfficientNet-V2, ViT, ConvNeXt)

--GPU access for deeper fine-tuning

--Learned segmentation models (U-Net, Mask R-CNN)

--Accuracy can be pushed well above 90%.

This project lays the foundation — the system is fully functional and can grow as better data becomes available.

6. System Architecture

User → Streamlit UI → Flask Backend → Segmentation → ResNet Model → Prediction → UI Visualization

Frontend (Streamlit)

--Upload up to 5 images

Displays:

--Annotated images

--Cropped cells

--Prediction labels

--Probability scores

--Backend (Flask)

--Accepts images

--Performs segmentation + preprocessing

--Runs inference

Saves:

--Annotated outputs

--Cropped cells

CSV logs

7. How to Run

Backend

--cd backend --pip install -r requirements.txt --python app.py

Frontend

--cd frontend --pip install -r requirements.txt --streamlit run streamlit_app.py

8. Folder Structure

project/
│
├── LDwebapp/
│   ├── backend/
│   │   ├── app.py
│   │   ├── leukemia_model.keras
│   │   ├── utils.py
│   │   ├── requirements.txt
│   │   └── static/
│   │       └── saved/
│   │
│   └── frontend/
│       ├── streamlit_app.py
│       ├── sample_input.jpg
│       └── requirements.txt
│
└── README.md

9. Disclaimer

This tool is created strictly for educational and research purposes.

It is NOT a clinical diagnostic tool and should not be used as a medical decision-making system.