Radiomics model and deep learning model based on T1WI image for acute lymphoblastic leukemia identification

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Acute lymphoblastic leukemia (ALL) diagnosis in children often relies on imaging.
• Developing accurate, non-invasive detection methods is crucial for early intervention.
• T1-weighted imaging (T1WI) offers potential for quantitative analysis.

Purpose of the Study:

• To develop and compare highly precise radiomics and deep learning models for ALL detection.
• To evaluate the diagnostic efficacy of these models using T1WI brain MRI.
• To assess the accuracy, sensitivity, and specificity of the developed models.

Main Methods:

• Utilized 604 T1WI brain MRI scans from children with ALL and healthy controls.
• Extracted radiomics features and trained a support vector machine model.
• Developed a deep learning model using the Efficientnet-B3 network.
• Validated both models on a separate testing cohort, calculating AUC, accuracy, sensitivity, and specificity.

Main Results:

• The deep learning model achieved a higher Area Under the Curve (AUC) of 0.981 compared to the radiomics model's 0.962.
• Deep learning model demonstrated superior accuracy (0.9344) and specificity (0.9737).
• Radiomics model showed higher sensitivity (0.9565) and negative predictive value (0.9714).

Conclusions:

• Both radiomics and deep learning models exhibit high diagnostic efficacy for ALL detection using T1WI MRI.
• Deep learning models show a slight advantage in overall performance, particularly in specificity.
• These AI-driven approaches hold promise for improving the accuracy and efficiency of ALL diagnosis.