Multi-Class Leukemia Detection Using a Lightweight Multipath Convolutional Neural Network

1. Introduction

      Leukemia is a type of blood cancer that affects the production and function of white blood cells. Early and accurate diagnosis is critical for effective treatment and improved survival rates. Traditional diagnosis methods rely heavily on manual examination of blood smear images by medical experts, which is time-consuming, subjective, and prone to human error.

      With the advancement of deep learning, automated image-based classification systems have gained significant attention in medical diagnosis. In this work, a Lightweight Multipath Convolutional Neural Network (LMCNN) is proposed for multi-class classification of leukemia using blood cell images. The model aims to provide an efficient and accurate solution suitable for real-world clinical assistance.

2. Problem Statement

Manual classification of leukemia subtypes from microscopic blood images is:

1. Time-consuming
2. Requires expert knowledge
3. Prone to inconsistency and errors

Additionally, existing deep learning models:

1. Are computationally expensive
2. Require large-scale resources
3. May not be suitable for lightweight deployment

Hence, there is a need for:

A lightweight, efficient, and accurate multi-class classification model for leukemia detection that can be easily deployed in real-world environments.

3. Proposed Method

This work proposes a Lightweight Multipath Convolutional Neural Network (LMCNN) for classifying blood cell images into four categories:

1. Benign
2. Malignant Pre-B
3. Malignant Pro-B
4. Malignant Early Pre-B

Key components:

1. Image preprocessing (resize + normalization)
2. Data balancing using oversampling with augmentation
3. Custom CNN architecture (LMCNN)
4. Performance evaluation using multiple metrics
5. Deployment using Django web application

The model is designed to maintain a balance between accuracy and computational efficiency.

4. Input and Output Specifications

Input

1. Blood cell microscopic images
2. Format: JPG/PNG
3. Size: Resized to 224 × 224 × 3 (RGB)

Output

1. Predicted class label:
2. Benign
3. Malignant Pre-B
4. Malignant Pro-B
5. Malignant Early Pre-B
6. Confidence score (%) of prediction

5. Methodology

The overall workflow of the proposed system is as follows:

Step 1: Data Collection

1. Blood cell image dataset containing four classes

Step 2: Data Preprocessing

1. Removal of invalid images
2. Conversion from BGR to RGB
3. Image resizing to 224×224
4. Normalization (pixel scaling)

Step 3: Data Balancing

1. Class imbalance handled using oversampling with augmentation
2. Techniques used:
3. Random flipping
4. Rotation
5. Brightness adjustment

Step 4: Model Design (LMCNN)

1. Three convolutional layers:
2. 32 filters → 64 filters → 128 filters
3. ReLU activation
4. Max pooling layers
5. Fully connected layers with dropout

Step 5: Model Training

1. Loss function: CrossEntropyLoss
2. Optimizer: Adam
3. Training with train-validation split
4. Evaluation on test dataset

Step 6: Performance Evaluation

1. Accuracy
2. Precision
3. Recall
4. F1-score
5. ROC-AUC
6. Confusion Matrix

Step 7: Deployment

1. Django-based web application
2. Image upload interface
3. Real-time prediction with confidence score

6. Novelty and Contribution  

The key contributions of this work include:

1. Development of a lightweight CNN model (LMCNN) for multi-class leukemia detection
2. Effective handling of class imbalance using data augmentation-based oversampling
3. Design of a computationally efficient architecture suitable for low-resource environments
4. Integration of model into a user-friendly Django web application
5. End-to-end pipeline from preprocessing to deployment

7. Expected Outcomes

The proposed system is expected to:

1. Accurately classify leukemia types from blood cell images
2. Reduce dependency on manual diagnosis
3. Provide fast and reliable predictions
4. Achieve good performance with low computational cost
5. Serve as a supportive tool for medical professionals

8. Conclusion

     This work presents a deep learning-based approach for multi-class leukemia detection using a Lightweight Multipath CNN. The model effectively classifies blood cell images while maintaining computational efficiency. By combining preprocessing, data balancing, and a custom CNN architecture, the system achieves reliable performance. The deployment using Django further enhances its usability as a real-time diagnostic support tool.

Future improvements may include:

1. Use of advanced architectures (ResNet, EfficientNet)
2. Larger datasets
3. Integration with clinical systems

Future improvements may include using larger datasets, advanced deep learning models, and improved preprocessing techniques to enhance the accuracy and reliability of the system.

9. References

1.Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet Classification with Deep CNNs.

2.He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition.

3.Litjens, G., et al. (2017). A Survey on Deep Learning in Medical Image Analysis.

4.Shorten, C., & Khoshgoftaar, T. (2019). A survey on Image Data Augmentation.

5.    WHO Reports on Leukemia Diagnosis and Classification