Unencrypted login | home

Program Information

BEST IN PHYSICS (JOINT IMAGING-THERAPY): A Radiomics Approach to Predict Local-Regional Failure for Advanced Head-And-Neck Cancer Using Pre-Treatment and Early Follow-Up CTs

X Wang

X Wang*, K Nie , S Sozio , A Khan , N Yue , S Kim , Rutgers-Cancer Institute of New Jersey, Rutgers-Robert Wood Johnson Medical, New Brunswick, NJ


TH-AB-201-1 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201

Purpose: To develop a radiomics approach for early prediction of local-regional failure for advanced head-and-neck cancer using pre-treatment and early follow-up CTs.

Methods: 35 consecutive patients (age: 58.6±11.4) with advanced head-and-neck cancer (stage III and IV) were included. All patients received concurrent chemoradiation to 70Gy and underwent two CT scans, one before and the other 2-3 weeks after the initial treatment. A total of 140 quantitative radiomics features, including morphology (as volume, compactness etc.), first-order histogram-based features (as skewness, kurtosis), second-order texture-based features (as Gray Level Co-occurrence Matrix (GLCM) textures, Run-length matrices) from pre-treatment CT, early follow-up CT and the associated changes, along with clinical information (as age, TNM staging) were extracted for each patient. Logistic regression with leave-one-out test was used for the prediction of local-regional failure.

Results: The median follow-up for all the patients was 8 months (range: 3-56 months) and 32.4% showed local-regional failure (LRF). Tumor volume was smaller in non-recurrence group compared to LRF group, though not statistically significant with an area under ROC curve (AUC) of 0.55. When combined with other clinical characteristics as TNM stage, and relative volume change, the AUC got improved to 0.61. Interestingly, the pre-treatment tumor compactness, which defined as the ratio of surface to the volume with a lower index indicating a smoother and spherical shape tumor and higher index for irregular shape with speculated margin, are significant higher in LRF group (p<0.02). The internal homogeneity index as measured by GLCM texture also showed significant differences between two groups. When combining quantitative imaging traits with clinical information, the AUC could be further improved to 0.81.

Conclusion: The results showed the feasibility of using radiomics approach to help with prediction of local-regional failure in advanced head-and-neck cancer. Future work will include additional dataset to further validate the robustness of this study.

Funding Support, Disclosures, and Conflict of Interest: Cancer Center Support Grant P30CA072720, RBHS Precision Medicine Pilot Grant

Add this talk to vcal | ical | Contact Email: