2161 - Artificial Intelligence-Based Assessment of Combined Dosiomic and Radiomic Features for Prediction of Metastatic Recurrence in Soft Tissue Sarcoma Treated with Preoperative Radiotherapy
Presenter(s)
L. Ubaldi1,2, M. Loi3, D. Greto3, A. Retico4, M. Banini5, G. Francolini3, M. Mangoni6,7, P. Bonomo8, G. Simontacchi3, N. Bertini7, I. Meattini9, S. Pallotta10, L. Livi3, and C. Talamonti11; 1Department of Experimental and Clinical Biomedical Sciences, University of Florence, Firenze, Italy, 2National Institute for Nuclear Physics (INFN), Florence Division, Firenze, Italy, 3Radiation Oncology Unit, Azienda Ospedaliero Universitaria Careggi, University of Florence, Florence, Italy, 4National Institute for Nuclear Physics (INFN), Pisa Division, Pisa, Italy, 5Department of Experimental Clinical and Biomedical Sciences “Mario Serio”, University of Florence, Florence, Italy, 6Department of Biomedical, Experimental and Clinical Sciences, University of Florence, Firenze, Italy, 7Radiation Oncology Unit, Azienda Ospedaliero Universitaria Careggi, University of Florence, Firenze, Italy, 8Radiation Oncology Unit, Azienda Ospedaliero-Universitaria Careggi, University of Florence, Florence, Italy, 9Department of Experimental and Clinical Biomedical Sciences “M. Serio”, University of Florence, Florence, Italy, 10University of Florence, Department of Biomedical, Experimental and Clinical Sciences “Mario Serio", Florence, Italy, 11Department of Biomedical, Experimental and Clinical Sciences “Mario Serio", University of Florence, Florence, Italy
Purpose/Objective(s): Preoperative radiotherapy (RT) followed by surgery is an effective modality of treatment for localized high-risk Soft Tissue Sarcoma (STS). However, approximately 30-50% patients will develop metastases, and despite availability of clinical nomograms, the role of adjuvant anthracycline-based chemotherapy in unselected patient is controversial. Integration of dose distribution data and post-treatment imaging may provide useful information on tumor biological behavior and response to preoperative treatment that may guide further management. The aim of our study is to develop a predictive model for metastatic relapse in STS treated with preoperative radiotherapy and surgery by integrating radiomic and dosiomic features.
Materials/Methods: Data from a consecutive cohort of STS patients treated with preoperative RT and surgery from 2011 to 2020 were retrospectively reviewed. All patients received preoperative radiotherapy to a dose of 50 Gy in 25 fractions followed by surgery after 4 weeks. Prior to surgery, a restaging contrast-enhanced MRI was performed to assess response to treatment. For all patients for whom available, simulation CT, dose distribution and restaging T1 gadolinium-injected MRI were analyzed. Radiomic and dosiomic features were extracted using the Python package PyRadiomics, which enabled the extraction of 107 features. These features were subsequently used to train and test a Random Forest-based classifier aimed at predicting the onset of metastases. The training, optimization, and testing of the algorithms were performed using a 5-fold Nested Cross-Validation, a strategy that allows for obtaining realistic performance estimates of Artificial Intelligence-based algorithms. The algorithms were implemented using the Python package Scikit-Learn.
Results: The cohort included 40 patients. Among them, 23 experienced metastatic relapse, with a median follow-up time of 42 months (range 3-96). AUC for CT, dose distribution and restaging T1 gadolinium-injected MRI was respectively 0.51, 0.71, and 0.59. By combining all the aforementioned features, AUC increased to 0.73.
Conclusion: Our model integrating data from simulation CT, dose distribution and T1 gadolinium-injected post-treatment MRI can effectively predict metastatic relapse in STS patient following preoperative RT and surgery and may serve as a useful tool to guide personalized treatment, for instance use of adjuvant chemotherapy or early detection of metastasis for locoregional treatment in oligometastatic clinical scenarios.