Blood Cancer Prediction Model Based on Deep Learning Technique

Written by: National Cancer Institute 

Summarized by: Janya Kumar

Abstract

Would you rely on an AI deep-learning model to identify cancer more quickly and accurately than your doctor? With technology advancing rapidly, this is no longer science fiction. As blood cancer continues to impact lives globally, early detection is essential for saving those affected. Blood cancer, also called hematological malignancy, starts in the bone marrow and affects red and white blood cells and platelets—the parts of blood that keep us alive and healthy. The most prevalent types include leukemia, lymphoma, and multiple myeloma. Symptoms can differ significantly and may consist of fatigue, frequent infections, and unusual bruising. Due to its high mortality rate, early diagnosis is crucial. This research investigates how deep learning techniques can revolutionize blood cancer detection. By employing advanced models such as ResNetRS50 and RegNetX016, the study demonstrates how artificial intelligence can achieve nearly perfect accuracy, transforming the way we diagnose and treat cancer. The findings indicate that ResNetRS50, in particular, delivers outstanding performance, offering a promising option for early intervention and improved patient outcomes.

Deep Learning in Cancer Diagnosis

The study reviews different deep learning models for identifying blood cancer: 

• ResNetRS50: 97% accuracy 

• RegNetX016: 96.6% accuracy 

• AlexNet: 87.9% accuracy

•  Convnext: 94.2% accuracy 

• EfficientNet: 93.0% accuracy

•  Inception_v3: 92.7% accuracy

•  Xception: 91% accuracy

• VGG19: 93.5% accuracy

ResNetRS50 surpassed all other models because of its deep structure, residual connections, and effectiveness. It was particularly good at detecting small yet significant features in medical images.

Methods

The research utilized the CNMC dataset, which includes more than 15,000 labeled leukemia

images. The images were resized to 224x224 pixels, denoised, and augmented to improve model generalization. 

ResNetRS50 and RegNetX016 were selected due to their established effectiveness in medical imaging. Both models were fine-tuned with extra custom layers, regularization, and dropout methods. The training of the models was conducted using the Adamax optimizer (with a learning rate of 0.001) for 40 epochs and a batch size of 40. ResNetRS50 delivered the highest performance across all evaluation metrics.

Results and Analysis

Immunotherapy helps the body’s own immune system recognize and destroy cancer cells. One of the most exciting developments is CAR T-cell therapy, which involves modifying a patient’s immune cells to enhance their ability to fight leukemia. This therapy has been approved for certain types of acute lymphoblastic leukemia (ALL) in children and young adults and is being studied for use in adults and other types of leukemia. 

Performance Metrics Table: 

Model	Accuracy	Recall	F1 Score
ResNetRS50	97%	99%	98%
RegNetX016	96.4%	99%	98%
AlexNet	87.9%	90%	88%
Convnext	94.2%	94.6%	95%
EfficientNet	93.0%	88%	91%
Inception_v3	92.7%	88%	91%
Xception	91%	89%	92%
VGG19	93.5%	93.6%	93.3%

Challenges and Future Directions

• Ethical Concerns: Need to address model bias

• Data Diversity: Expand datasets to include various demographics

• Clinical Trials: Further validation in real-world settings

• Interpretable AI: Develop explainable models for better clinical trust

• Global Accessibility: Adapt for use in low-resource settings

Conclusion

ResNetRS50 stands out as the most effective model for blood cancer detection with a 97% accuracy rate. Its robust architecture and efficiency make it ideal for integration into clinical systems, potentially transforming early cancer diagnosis and treatment worldwide.

 Works Cited

National Cancer Institute. “Advances in Leukemia Research.” National Cancer Institute, Cancer.gov, 25 June 2019, www.cancer.gov/types/leukemia/research.