5506. Radiomics Risk Prediction Model to Predict Clean Postsurgical Tissue Margin in Soft Tissue Sarcoma
Authors * Denotes Presenting Author
  1. Ehsan Alipour *; University of Washington
  2. Atefe Pooyan; University of Washington
  3. Arash Azhideh; University of Washington
  4. Firoozeh Shomal Zadeh; University of Washington
  5. Chankue Park; University of Washington
  6. Matthew Nyflot; University of Washington
  7. Majid Chalian; University of Washington
Soft tissue sarcoma tumors require precise planning for neoadjuvant radiotherapy and surgical resection to achieve local control. One of the most important determinants of a high chance of local recurrence in these tumors is the involvement of surgical margin with tumoral tissue. In this study, we propose a multimodal to predict clean postsurgical margin using pretreatment MRI radiomics features and clinical information about the tumor. The model can be used to flag patients who are at a high risk of having margin involvement following surgery.

Materials and Methods:
We identified a retrospective cohort of patients with soft tissue sarcoma in our institute. Selection criteria included availability of pretreatment MRI imaging and a postresection pathology report that contained information about the margins of the removed tumor. Additionally, clinical information including sex, age, tumor location and grading were extracted. Two musculoskeletal radiologists with 10 years of experience manually segmented out tumor regions on the pretreatment MRIs on four sequences including fat saturated T2, ffat saturated T1 before and after intravenous contrast and T1. A total of 104 original radiomics features were extracted from each sequence using the Pyradiomics package (416 total features from 4 sequences for each patient). Early fusion was used to pool all the data together. Margin involvement information was manually extracted from the postsurgical pathology reports. Data were divided into training sets and testing sets using a 0.8,0.2 ratio. A XGBoost model was trained to classify patients based on the outcome. Ten-fold cross validation was used to select the best model. Model calibration was performed based on the probabilities in the train set, and the best threshold for model prediction was selected using the Youden index method. Final model was tested on the holdout set and Shapley values were used to assess feature importance and explain the model’s performance.

We identified 199 patients with soft tissue sarcoma that fit our selection criteria were identified. Average age of patients at the time of diagnosis was 54 years old. There were 70 women and 129 men patients. Out of these, 149 had a clear tissue margin. Our best performing model achieved a ROC-AUC of 0.73 on the test set reaching a sensitivity of 60% and specificity of 73% using the threshold suggested by the Youden index. The negative predictive value of the model was 85%. The final model only used 10 of the input features. Most important features were original shape sphericity of the T1AC image and tumor’s maximum 2D diameter.

Our multimodal radiomics-based risk model can be used to predict the involvement of postsurgical tissue margin in STS. This information can be used in routine clinical practice to inform clinicians so that they can adjust their treatment plan. In addition, since the model is explainable and uses a small number of features, it can be used to devise a simplified criteria to identify patients who are at a higher risk for postsurgical margin involvement.