ESTRO 2024 - Abstract Book

S5004

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Twenty-eight LARC patients who received a short course (25Gy, 5 x 5Gy/fr) of MRgART using a 1.5 Tesla MR-Linac, followed after one week by chemotherapy, were included in this retrospective study. Patients after nCRT were evaluated if cCR or not; if patients resulted cCR, they followed a watch-and-wait strategy with regular restanging by MRI and digital rectal examination; in the absence of cCR the surgery was scheduled then pCR was evaluated following the classification system of Mandard et al. [4]; For each patient, 3D-T2-weighted transversal MRI images acquired during the online adaptive clinical workflow for each of the five adaptive fractions (fractions 1 to 5, f1 to f5)[5], corresponding target delineation, pCR result of the surgical specimen and clinical variables were collected. The gross tumor volume obtained in the clinical practice was used for radiomics study and the analysis was focused on the whole GTV for each of the five fractions. MRI images were normalised using Z-score methods, the Laplacian of Gaussian (LoG) filters were applied to the raw MRI images; LoG filters using a range of standard deviations from 0.2 to 1.0 mm with step size 0.2 mm were used. Seven families of features (First order, Shape, GLCM, GLDM, GLRLM, GLSZM and NGTDM) were extracted and delta features were calculated from the ratio of features of each successive fraction to those of the first fraction. Wilcoxon-Mann-Whitney test and LASSO were utilized to reduce the dimension of features and select those features that are most significant to pCR. At last, the radiomics signatures were established by linear regression with the final set of features and their coefficients. Among the 28 patients enrolled, 26 (92.9%) were identified as cCR patients and underwent TME surgery at the end of chemotherapy. Of the two cCR patients (7.1%), one remained cCR for 22 weeks after the end of nCRT and then tumour recurrence was detected and underwent surgery; the second, 45 weeks after the end of nCRT is still cCR to date. By the end, 27 patients had already undergone surgery and histopathological examination revealed that 5 of them (18.5%) were pCR. A total of 581 radiomics features were extracted and 2324 delta features were calculated from the 581 features of each of the four fractions following the first. The Wilcoxon-Mann-Whitney test revealed that nineteen features and delta features, and one clinical variable (cN) were significant (p < 0.05) to pCR. Using LASSO regression, the dimension of the 20 variables was further reduced towards zero, thus retaining only the most relevant features and discarding the less important ones. The optimal lambda value (l.min) of 0.0245, was obtained when the LASSO regression achieved the lowest binomial deviance using LOOCV validation. With this optimal lambda value, 13 features had a zero coefficient in the Lasso regression and were eliminated; of the 7 remaining, one clinical variable (cN), one feature from fraction 1, and 6 delta features . The 7 variables and their corresponding coefficients, shown in Table 1, were used to construct by LASSO regression the Rad-Score for each patient . Figure 1 shows that the Rad Score can significantly discriminate pCR and non-pCR patients (p < 0.05). Results:

Table 1. The optimal set of features and corresponding coefficients used to construct the rad score for patients