Manifold Learning-based Data Sampling for Model Training

Zusammenfassung

Training data sampling is an important task in machine learning especially for data with small sample size and data with nonuniform sample distribution. Dividing data into different data sets randomly can cause the problem that, the training model covers only parts of the sampled cases and works inaccurately for weakly sampled cases. Recent research showed the benefit of manifold learning techniques in medical image processing. In this work, we propose a manifold learning based approach to improve the data division and the model training. We evaluated the proposed approach using an atlas registration framework and a deep learning framework. The final segmentation results using methods with and without data balancing were compared. All of the final segmentations were improved after implementing the manifold learning based approach into the frameworks. The largest improvement was 24.4%. Thus, the proposed manifold learning based approach is effective for the model training.