ESTRO 2024 - Abstract Book

S4403

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

is ~0.84, only slightly worse than the median F1 measure of ~0.85 achieved by our proposed TL-DSN-JC. This indicates that our committee method using the classic VGG model still holds up against the transformer-based SAM with a small but clear edge. The results predicted by the TL-DSN-JC algorithm were evaluated using three performance measures: precision, recall, and the harmonic mean of them denoted by F1, whose average values were found to be (0.97, 0.93, 0.95) for typical cases. The TL-DSN-JC algorithm outperformed other similar algorithms such as the singular DSN without transfer learning, DSN jury committee without transfer learning, and singular DSN with transfer learning by up to 80%.

Figure 2. Precision (left), Recall (middle), and F1 (right) measures for four methods using jury committee (JC) and transfer learning (TL): nTL-single-DSN: singular DSN without Transfer Learning; nTL-DSN-JC: DSN jury committee without Transfer learning; TL-single-DSN: singular DSN with transfer learning; SAM: Meta’s Segment Anything Model; TL-DSN-JC: DSN jury committee with transfer learning, the actual method proposed in this work.

Conclusion:

In this work, a novel transfer learning based deep segmentation net with jury committee algorithm was developed to extract the mask of the lung tumor target from the low contrast and noisy kV radiographs. Our results demonstrated that the proposed algorithm could reliably segment the lung tumor on these images and outperformed the conventional deep learning techniques by up to 80%. Our new algorithm has a great potential to detect the lung tumor accurately based on sequential kV projection images during the radiotherapy of lung cancer.

Keywords: Tumor motion tracking; transfer learning

3188

Digital Poster

Dosimetric assessment of intra-fraction motion during online adaptive MRI-guided RT