Introduction of Research on Automatic Early-Stage Lung Cancer Segmentation Using Fractal Images and Its Potential Application to AiRato’s Technology

AiRato Inc. (Head Office: Sendai, Miyagi; CEO: Yuri Kimura), a company developing and commercializing AI-based radiation therapy planning support services, is paying close attention to a new transfer learning framework announced by research teams from the Department of Radiation Oncology, Tohoku University Graduate School of Medicine, and the University of Yamanashi.

In this study, the researchers report that self-supervised learning using mathematically generated fractal images achieved more accurate automatic segmentation of the Gross Tumor Volume (GTV) in early-stage lung cancer than conventional ImageNet-pretrained models.

According to the paper, when the Vision Transformer (ViT) model was pretrained on FractalDB-10K, it achieved a vDSC of 0.800 ± 0.079 and an HD95 of 2.04 ± 1.59 mm, significantly outperforming the ImageNet-1K-pretrained model (vDSC = 0.764 ± 0.102; HD95 = 3.03 ± 3.47 mm).

The improvement was particularly notable in cases with complex tumor shapes, which is thought to be due to the self-similarity of fractal images—an attribute they share with tumors themselves.

AiRato believes that these findings can be applied to enhance the accuracy and learning efficiency of its own automatic contour extraction algorithms used in its radiation therapy planning support system currently under development.

By incorporating the techniques and insights demonstrated in this joint research by Tohoku University and the University of Yamanashi, AiRato aims to deliver more precise and efficient treatment planning solutions.

Moving forward, AiRato will continue to actively adopt cutting-edge academic research and technological trends to further advance its AI solutions for radiation therapy—ensuring that both patients and healthcare professionals can use them with confidence and peace of mind.

Publication Information

• Title: Fractal-driven self-supervised learning enhances early-stage lung cancer GTV segmentation: a novel transfer learning framework
• Authors: Ryota Tozuka, Noriyuki Kadoya, Arata Yasunaga, Masahide Saito, Takafumi Komiyama, Hikaru Nemoto, Hidetoshi Ando, Hiroshi Onishi, Keiichi Jingu
• Journal: Japanese Journal of Radiology (Original Article)
• Link: https://doi.org/10.1007/s11604-025-01865-8