Room: Exhibit Hall | Forum 5
Purpose: This work aim is to investigate if poor survival and early recurrence may be associated with similar dose distribution patterns delivered to the organ with tumor. On the example of metastatic liver cancer, we harnessed the power of convolutional neural networks to automatically identify such dose patterns.
Methods: A database of 58 SBRT cases for metastatic liver cancers with post-treatment follow-ups has been collected. Each case consisted of the CT image with patientâ€™s anatomy segmented and the executed 3D dose plan. The survival records were converted into binary values where the 2-year cut-off points separated positive and negative survival. Similarly, the local progressions were converted into binary values with 1-year cut-off separator. We designed two 3D deep convolutional neural networks (CNNs) to associate 3D dose plans with different treatment outcomes. The behavior of CNNs was analyzed to understand irradiation of which liver regions is considered to impose the highest risks of negative outcome. Finally, we estimated how statistically similar are the dose patterns associated with high risks of both outcomes.
Results: The risk scores generated by CNNs were accumulated for eight individual liver segments. A high risk score indicate that the irradiation of the corresponding segments imposes high risks of a negative SBRT outcome. We observed that the highest risks of both negative outcomes was associated with irradiation of the Ist liver segment. We attribute it to the fact that critical liver vessels pass in or near the Ist liver segment. Risk scores for all segments computed for survival and local progression significantly correlate with coefficient 0.85.
Conclusion: A novel CNN-based framework for mapping dose distribution patterns and outcomes after metastatic liver SBRTs has been developed. Our results confirms the assumption that similar dose plans may lead to both early post-treatment death and local cancer recurrence.
Dose Response, Radiation Therapy