ESTRO 2024 - Abstract Book

S3992

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

Results:

The 6-layer CNN architecture outperformed all other architectures among projection views. The overall best performing model was the dual-view late feature fusion 6-layer CNN, with the lowest MAE of (66±9) mL and the highest robustness across seeds (ranging between [71±18, 192±229] mL in other networks). The dual-view ensemble prediction models proved to be the most resilient to architecture changes, with dual-view ensemble 6-layer CNN exhibiting the lowest MAPE of (49±5) % (other networks: [52±11, 158±186] %). Lateral models were superior to frontal models, both in terms of accuracy and capturing linear trends in data (r lat ∈ [0.87, 0.90] vs. r front ∈ [0.51, 0.71]), however, methods that employed both views showed superior performance. Fig. 2 shows the absolute differences between the expected and ensemble GI gas volume predictions obtained from applying the five dual-view late feature fusion 6-layer CNN models on the held-out TCIA and independent clinical test sets. The predictive performance achieved on the clinical dataset was superior to that on the TCIA dataset, both in terms of MAE (22 vs. 57 mL) and MAPE (24 vs. 48 %). GI gas volume predictions on both datasets were strongly correlated to the reference values (r TCIA = 0.90 and r clinical = 0.89).

Conclusion:

This work demonstrated the feasibility of evaluating internal anatomical change from planar radiographs. Our method may inform the need for volumetric imaging while minimising diagnostic exposure in young patients who are at greater risk of developing second cancers later in life. Shallower networks using multiple input projections generated the most accurate and reliable predictions. Initial findings on a limited independent dataset hint at the promising generalisation ability of the models, laying the groundwork for more extensive future evaluations. This