X Chen¹*, M Zhou¹, K Wang², Z Wang⁴, Z Zhou⁴, (1) Xi'an Jiaotong University, Xi'an, Shaanxi, CN, (2) UT Southwestern Medical Center, Dallas, TX, (3) Peking University Cancer Hospital, Beijing, CN, (4) University Of Central Missouri, Warrensburg, Missouri

• 
  X Chen

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
The aim of this study is to predict immunotherapy response (progress or pseudoprogress) in metastatic melanoma by developing a new reliable automated multi-objective delta-radiomic (Auto-MODR) model.

Methods:
In this study, totally 50 patients (Peking University Cancer Hospital, Beijing, China) who underwent contrast-enhanced computed tomography (CECT) examinations pre and one-cycle post-treatment were used. All these patients were diagnosed pathologically with metastatic melanoma. After extracting radiomic features (including intensity, texture, and geometry features) from pre-treatment and one-cycle post-treatment CECT images respectively, we obtained delta-radiomic features by calculating the difference between pre-treatment radiomic features and one-cycle post-treatment radiomic features. Then the different combinations of pre-treatment radiomic features, one-cycle post-treatment radiomic features and delta-radiomic features were used to predict the future immunotherapy response. An automated multi-objective delta-radiomic (Auto-MODR) model was developed to predict immunotherapy response in this study. Auto-MODR incorporated delta-radiomic features with traditional radiomic features to generate the Pareto-optimal model set through multi-objective optimization. To take advantage of all generated Pareto-optimal models set, the evidential reasoning (ER) strategy was used to fuse the output probabilities of these models. The label with the maximal output probability is considered as the final label output.

Results:
Combining traditional radiomic features with delta-radiomic features outperformed the model with traditional radiomic features only. Particularly, the model combining one-cycle post-treatment radiomic features and delta-radiomic features had a highest AUC of 0.829, whereas the model that just used one-cycle post-treatment radiomic features only yielded AUC of 0.728. Furthermore, Auto-MODR outperformed traditional multi-objective model (MO) and traditional single-objective model (SO-AUC).

Conclusion:
We developed a new automated multi-objective delta-radiomic (Auto-MODR) model for predicting immunotherapy response (progress or pseudoprogress) in metastatic melanoma. The experimental result demonstrated that the best performance can be obtained when combining the traditional radiomic features with delta-radiomic features.

Keywords

Feature Selection, Image Analysis, Modeling

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email