Abstract
BACKGROUND: Manual segmentation of brain tumor images from a large volume of MR images generated in clinical routines is difficult and time-consuming. Hence, it is imperative to develop a machine learning model for automated segmentation of brain tumor images. PURPOSE: Machine learning models for automated MR image segmentation of gliomas may be useful. However, the image differences among facilities cause performance degradation and impede successful automatic segmentation. In this study, we proposed a method to solve this issue. METHODS: We used the data from the Multimodal Brain Tumor Image Segmentation Benchmark (BraTS) and the Japanese cohort (JC) datasets collected from 10 facilities. Three models for tumor segmentation were developed. The BraTS model was trained on the BraTS dataset, and the JC model was trained on the JC dataset; whereas, the Fine-tuning model was a fine-tuned BraTS model using the JC dataset. RESULTS: MR images of 544 patients were obtained for the JC dataset. Half of the JC dataset was used for independent testing. The Dice coefficient score of the JC model for the JC dataset was 0.779± 0.137, whereas that of the BraTS model was remarkably lower (0.717 ± 0.207). The mean of the Fine-tuning models for the JC dataset was 0.769 ± 0.138. There was a significant difference between the BraTS and JC models (P < 0.0001) and the BraTS and Fine-tuning models (P = 0.002); however, no significant difference was observed between the JC and Fine-tuning models (P = 0.673). CONCLUSIONS: Application of the BraTS model to heterogeneous datasets can significantly reduce its performance; however, fine-tuning can solve this issue. Since our fine-tuning method only requires less than 20 cases, this methodology is particularly useful for a facility where there are a few glioma cases.