Automatic head and neck tumor segmentation through deep learning and Bayesian optimization on three-dimensional...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computational pathology

Background:

• Medical imaging is vital for diagnosing diseases, relying heavily on clinician expertise.
• Convolutional Neural Networks (CNNs) show promise for medical image analysis but face hyperparameter optimization challenges.
• Optimizing hyperparameters is crucial for effective CNN implementation in clinical settings.

Purpose of the Study:

• To propose a novel two-phase Bayesian Optimization-derived Scheduling (BOS) approach for hyperparameter optimization in medical image segmentation.
• To enhance the accuracy and efficiency of CNNs in identifying pathologies and segmenting tissues.
• To address the challenge of optimizing numerous hyperparameters in CNNs for clinical applications.

Main Methods:

• A two-phase Bayesian Optimization-derived Scheduling (BOS) approach was developed for hyperparameter optimization.
• Batch size and learning rate were coupled as a batch size to learning rate (B2L) ratio for simultaneous optimization.
• The optimized hyperparameters were applied to a 3D V-Net model using CT and PET scans for tissue segmentation.

Main Results:

• The two-phase BOS approach demonstrated improved performance in head and cancerous tissue segmentation tasks.
• Coupling batch size and learning rate (B2L ratio) led to more optimal hyperparameter combinations.
• 10-fold cross-validation indicated enhanced overall medical image segmentation performance with the optimized B2L ratio.

Conclusions:

• The proposed two-phase BOS method effectively optimizes CNN hyperparameters for medical image segmentation.
• Simultaneous optimization of batch size and learning rate via the B2L ratio is beneficial.
• This approach holds potential for improving diagnostic and prognostic evaluations in medical imaging.