ESTRO 2024 - Abstract Book

S3013

Physics - Autosegmentation

ESTRO 2024

volume-level sensitivity-specificity (VSS) loss [5] and the binary cross entropy (BCE) loss are used to improve metastases detection and segmentation, respectively. The DeepMedic network was used as it effectively addresses the class imbalance problem by random sampling of subvolumes. As a starting point of multicenter collaboration, this work focuses on bilateral collaboration with a model pretrained in Erlangen and being shared to Zurich or BraTS.

Results:

1. Erlangen model shared to Zurich: When the models were independently trained, the Erlangen model achieved 0.8419 sensitivity and 0.9347 precision (metastasis-wise instead of voxel-wise) on Erlangen data, and the Zurich model achieved 0.8611 sensitivity and 0.7561 precision with brain mask (or 0.3360 precision without brain mask) on Zurich data (see Table 2). As the Zurich dataset includes metastases of leptomeningeal origin near cortical surfaces and meninges, the trained model predicted many false positive (FP) metastases (205 FP metastases out of in total 245 FP metastases) outside the brain region. Such FP metastases can be simply removed with binary brain masks. This behavior is only observed in the Zurich dataset, but serves as a good indicator of model performance and knowledge transferability. When the Erlangen pretrained model was directly applied to Zurich test data without further fine tuning, the model achieved 0.8194 sensitivity and 0.7516 precision without brain mask. Even if no brain masks were used, no FP metastases were predicted outside the brain region. This demonstrates the good generalizability of the Erlangen model, due to its relatively large training dataset. When the Erlangen model was fine-tuned without LWF on Zurich training dataset for 150 epochs, the fine-tuned model achieved 0.8681 sensitivity and 0.5230 precision with brain mask (0.2969 precision without brain mask) for Zurich test data, and it achieved 0.8971 sensitivity and 0.2255 precision with brain mask (0.1773 precision without brain mask) on Erlangen test data. These values indicate that the fine-tuned model overfits to Zurich data and forgets the knowledge learned from Erlangen data. However, when LWF was used for fine-tuning, the LWF model achieved 0.8472 sensitivity and 0.7485 precision without brain mask for Zurich test data. No FP metastases were predicted outside the brain region, indicating the benefit of LWF in the target center. The LWF model achieved 0.8603 sensitivity and 0.8897 precision on Erlangen test data, demonstrating the learned knowledge from Erlangen training dataset was preserved. 2. Erlangen model shared to BraTS: The model trained on BraTS data independently achieved 0.8498 sensitivity and 0.4776 precision on BraTS test data. More cases with lower image quality (e.g., noise, motion artifacts, and low resolution) were observed in the BraTS data than the Erlangen data, which leads to a higher FP rate. The fined-tuned model without LWF achieved 0.8826 sensitivity and 0.4563 precision on BraTS test data. It also achieved a high sensitivity of 0.9044 on Erlangen test data, but with a lower precision of 0.6814. Fine-tuning with LWF achieved 0.8498 sensitivity and 0.4801 precision on BraTS test data, and it achieved 0.8419 sensitivity and 0.9308 precision on Erlangen test data, where the sensitivity and precision were preserved.