Model Card for FFSCS (Anterior Mediastinal Tumor Segmentation)

This model is a 3D U-Net-based deep learning system designed for the automated segmentation and detection of anterior mediastinal tumors on chest CT images. It was developed to overcome the challenges of rare disease diagnosis using a multi-institutional dataset from 136 hospitals.

Model Details

• Developed by: Chihiro Takemura, Hirokazu Watanabe, et al. (National Cancer Center Hospital, Japan)
• Model type: 3D U-Net (Segmentation)
• Language(s): English (Documentation)
• License: Apache-2.0
• Platform: Developed using SYNAPSE Creative Space (No-code AI platform by FUJIFILM Corporation)

Intended Use

• Primary Use: Segmentation and detection of anterior mediastinal tumors (e.g., Thymoma, Thymic Carcinoma, Lymphoma, Thymic Cyst) in 3D CT scans.
• Intended Users: Medical AI researchers, diagnostic radiologists, and thoracic surgeons.
• Out of Scope: This model is for research purposes only and is not cleared for primary clinical diagnosis or use as a medical device.

Technical Specifications

• Architecture: 3D U-Net
• Input: 3D CT volumes (normalized to [0, 1]).
• Output: Binary segmentation mask (0: Background, 1: Tumor).
• Training Hyperparameters: - Optimizer: Adam (Learning rate: 0.001)
• Loss Function: Dice Loss
• Batch size: 3
• Data Augmentation: Random rotation, scaling, and cropping.

Contact

For questions regarding the model, please contact the corresponding author: Hirokazu Watanabe ([email protected]).