Optimization of brain MRI preprocessing using parallel computing for efficient brain tumour classification.
Saved in:
| Title: | Optimization of brain MRI preprocessing using parallel computing for efficient brain tumour classification. |
|---|---|
| Authors: | Mochurad, Lesia1 (AUTHOR) lesiamochurad@gmail.com, Melnychuk, Khrystyna1 (AUTHOR) khrystyna.melnychuk.kn.2021@lpnu.ua, Mochurad, Yulianna2 (AUTHOR) yuliannamochurad@gmail.ua |
| Source: | Journal of Supercomputing. Feb2026, Vol. 82 Issue 3, p1-31. 31p. |
| Abstract: | In medical informatics, efficient preprocessing is essential for convolutional neural network (CNN)-based analysis of brain magnetic resonance imaging (MRI), particularly for image-level tumour classification. One of the main challenges is the large volume of medical data, which requires substantial computing resources for preprocessing. In this study, we focus on a four-class brain tumour classification task and aim to optimise the MRI image preprocessing stage using parallel computing. In particular, we employ the map function from the TensorFlow Dataset API and the concurrent.futures.ThreadPoolExecutor module to substantially reduce the processing time of large datasets used to train CNNs. The results show that, for offline dataset materialisation (Stage 1), the proposed workflow accelerates preprocessing by 27.2 × compared with a naïve prototype script and by 2.85 × compared with a standard TensorFlow baseline. For the per-epoch input pipeline (Stage 2), tuning num_parallel_calls provides an additional 1.63 × speedup over the standard baseline. The number of threads was also tuned empirically, providing additional acceleration, with the best results obtained when using 4 threads for the map function and 8 threads for the ThreadPoolExecutor. In addition, we investigated the effect of image-processing filters; in a single controlled run, the highest validation accuracy observed was 94.58% with the blur filter. The proposed approach improves medical image preprocessing through parallel computing, which substantially reduces processing time and increases the efficiency of CNN-based brain tumour classification. The same preprocessing and parallelisation strategy is largely task-agnostic at the data-input level and may serve as a building block in more complex pipelines; however, segmentation-specific evaluation is left for future work. [ABSTRACT FROM AUTHOR] |
| Database: | Academic Search Index |
Full Text Finder 
Nájsť tento článok vo Web of Science Be the first to leave a comment!