| 日 時 | 2026年11月24日(火) 15:00 より 15:50 まで |
|---|---|
| 講演者 | Rizal Dwi Prayogo, Ph.D. |
| 場 所 | 生理学研究所 １Fセミナー室A/B |
| お問い合わせ先 | 福永雅喜（fuku@nips.ac.jp） |
| 要旨 |
Magnetic resonance imaging (MRI) provides detailed views of brain tissue that support tumor assessment. However, automated MRI-based brain tumor classification remains challenging because labeled data are limited, tumor appearances vary, and complex models require substantial computation. This seminar presents three studies. First, transfer learning adapted a model trained on general images to classify 253 brain MRI images. Data augmentation created controlled variations of training images, while regularization helped limit overfitting, where a model learns training-specific patterns that may not generalize to unseen images. This study examined binary tumor detection, leaving open how to represent varied patterns across multiple tumor categories. The next study combined features from three pretrained networks and achieved 98.75% accuracy on a 1,200-image test set. However, its 33.139 million parameters motivated a more compact architecture. HayabusaNet was designed to capture patterns at multiple spatial scales and integrate local image details with broader context. It achieved 98.86% accuracy on a 1,405-image validation set using 1.319 million parameters. External testing assessed performance across data sources, while visual explanation maps supported qualitative inspection of predictions. The evidence remains limited to two-dimensional image classification rather than patient-level or clinical validation. The seminar concludes with implications for future MRI analysis at NIPS. Keywords: Magnetic Resonance Imaging (MRI); Brain Tumor Classification; Deep Learning; Transfer Learning; Computational Efficiency. |
