Hyperparameter selection for dataset-constrained semantic segmentation: Practical machine learning optimization.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Optimizing machine learning models, especially with limited data, is challenging.
• Hyperparameter selection significantly impacts model performance in medical image analysis.

Purpose of the Study:

• To provide a pedagogical example of systematic machine learning optimization for small dataset image segmentation.
• To emphasize the importance of hyperparameter selection in medical image segmentation models.
• To present a simple, applicable process for medical physicists to examine hyperparameter optimization.

Main Methods:

• Developed a multiclass segmentation model using a public Computed Tomography (CT) dataset.
• Conducted preliminary manual hyperparameter search followed by grid search.
• Trained 658 models on 13,160 patients and analyzed results using random forest regression for hyperparameter impact.

Main Results:

• Discrepancy observed between metric-implied segmentation quality (e.g., 96.8% accuracy) and visual inspection.
• Batch normalization identified as a key hyperparameter, though performance varied.
• Grid search and random forest analysis proved to be an easily implementable sensitivity analysis approach.

Conclusions:

• The proposed method offers a systematic, quantitative approach to understanding hyperparameter influence on model performance.
• Grid search with random forest analysis provides valuable insights within hardware and data constraints.
• This methodology enhances model validity and reduces decision-making risks for medical physicists.