3651 - Tracking Radiotherapy Treatment Response through Time Using Deep Learning Autocontouring on CBCTs

Purpose/Objective(s): Radiation treatment dosages are one-size-fits-all. This lack of personalization is likely leading to over- and under-treatment. This work is a proof of concept to determine if the continuous auto-contouring of a gross tumor volume (GTV) through treatment can be used to predict a patient’s long-term outcome.

Materials/Methods: In prior work, we trained an auto-contouring model to segment the GTV for HNC using the Cancer Imaging Archive HN1 Dataset. We used this model with no re-training to contour the GTV for both the planning CT and subsequent CBCT scans to track changes in tumor size through treatment. We then used an unsupervised algorithm, Time-Series K-Means with DTW, to group the patients into three clusters with similarly shaped time-series, e.g. patients with a sudden drop in tumor size at the beginning of radiotherapy would be clustered into a different group than patients whose tumor size increased through time. Finally, we used a small amount of labeled data (six patients) to gain insight into whether the clusters were meaningful.

Results: Clusters 1 and 3 show a profile in which the GTV size dropped through radiotherapy, while Cluster 2 shows a profile in which the GTV size did not decrease. When we matched these three clusters with the labeled patients, we found that the NED/Improvement patients were assigned to Clusters 1 and 3 and the Persistent Disease patient was assigned to Cluster 2. About 15% of patients (50/335) were assigned to Cluster 2, which closely aligns with a study that indicates the cumulative rate of disease recurrence after one year is 18.6%.

Conclusion: There is a clear difference in the shape profile between Improvement/NED vs. Persistent Disease. This is proof of concept that we can use auto-contouring of CBCTs to track tumor volume through radiotherapy treatment, which enables us to predict who will have a positive response at 3 months, which is indicative of long-term outcomes. With further study, our goal is to identify if a patient is a candidate for (de-)escalation.

This method can also be used to address a widespread labeling problem in medicine; we enable researchers to determine if a strong enough signal exists in the data to invest the time and money to label the full dataset for supervised learning methods.

This work is a proof of concept that will lead to multiple benefits to society, including treatment selection, improving disease control and survival from HNC, while enhancing long-term quality of life.