Author: Mingli Chen, Xuejun Gu, Hao Jiang, Mahdieh Kazemimoghadam, Weiguo Lu, Qingying Wang, Kangning Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose:
This work demonstrates how existing software, when creatively adapted, can address a wide range of clinical challenges. By focusing on data exploration and application-specific modifications, we extend the capabilities of a segmentation software to tasks such as image synthesis, dose prediction, anomaly detection, and contour quality assurance, highlighting the potential of leveraging existing tools to solve diverse clinical needs efficiently and cost-effectively.
Methods:
We adapted the nnU-Net framework with minimal modifications to tackle four distinct applications. For MR-to-synthetic-CT-conversion, we replaced the loss function with Huber loss and make it an image-to-image-transformer, and trained models on registered MR-CT pairs. For PET-to-synthetic CT generation, a similar approach was applied to PET/CT datasets. For universal dose prediction, we use three input channels, including prescription, avoidance, and beam trace images, and trained it on a large dataset spanning 25 treatment sites. For anomaly detection, we converted nnU-Net into an auto-encoder to reconstruct segmentation masks and detect anomalies. Each adaptation required only minor changes, showcasing the framework’s flexibility.
Results:
The universal MR-to-synthetic-CT models work for any body sites of various sequences with high gamma passing rates (98.6% for 2%/2mm), demonstrating accuracy for MR-only planning. The PET-to-synthetic-CT model replicated anatomical structures without metal artifacts, offering a promising CT-free PET imaging. The dose prediction model achieved a 92.36% gamma passing rate, showing strong consistency with optimized plans. The anomaly detection model exhibited negligible errors for normal masks but significantly higher errors for anomalies, enabling reliable quality assurance.
Conclusion:
This work highlights the value of exploring existing frameworks for new data and applications, rather than “inventing” new architectures. By creatively adapting tools like nnU-Net, we can address diverse clinical challenges efficiently. This "No-New-Network" approach underscores the untapped potential of existing frameworks, offering a pathway to safer, more efficient, and innovative medical image and radiotherapy solutions.