The electronic medical database contained data for 105 rectal adenocarcinoma patients (adenocarcinoma < 10 cm from rectal lower margin) who underwent the standard long-course NCRT followed by radical resection between October 2017 and August 2022 at The First Affiliated Hospital of Hebei North University and The Fourth Hospital of Hebei Medical University. The inclusion criteria were as follows: (1) histologically diagnosed with primary rectal adenocarcinoma; (2) with locally advanced rectal cancer based on enhanced chest, abdominal, and pelvic CT, rectal MRI, and transrectal ultrasound, according to the eighth edition of the Joint Cancer Board (AJCC) (32) before treatment; (3) receiving neoadjuvant chemoradiotherapy and TME; (4) with preoperative MRI data. The exclusion criteria were as follows: (1) with incomplete standard NCRT. Eight patients did not complete the standard NCRT due to intolerance and rejection; (2) with other malignancies (six); (3) with no tumor regression grading data (four); (4) with low-quality key MRI images for analysis (three). In the end, 84 patients met the inclusion criteria. Further examination revealed that postoperative imaging data for 7 patients were missing. Finally, 84 patients with pre-Radiomics of MRI and 77 patients with delta-Radiomics MRI were included in this study ( Figure 1 ).

According to the time of data collection, we used the last 16 data collected as an independent validation cohort, and the other data as the primary cohort for model construction and cross-validation. Finally, the established pre-Radiomics and delta-Radiomics models were evaluated again with the validation cohort.

Patients with LARC received long-course radiotherapy (45–50Gy, 1.8–2.0 Gy each time, 5 times per week) and concomitant chemotherapy (capecitabine 1250 mg/m2 twice a day, 5 times per week). Radical excision (TME) was performed within 8-12 weeks after the completion of NCRT.

Postoperative TNM restaging was performed according to the pathological outcomes of the surgically resected specimens to evaluate the down-staging. Tumor regression response was evaluated systematically according to tumor regression grade (TRG) (32). The details were as follows: Grade 0: the tumor completely retracted, and only calcium salt deposition in the fibrous tissue showed pathological response; Grade 1,: moderate retraction. Here, fibrosis was present with a few visible tumor cells or cell masses; Grade 2: slight retraction. Here, there was no residual tumor, but strong fibrosis interstitial filling was present; Grade 3: no regression, extensive residual tumor, and little or no tumor cell necrosis. TRG0-1 was defined as a good reaction, whereas TRG2-3 was defined as a poor reaction.

In this study, all MRI image data were acquired from two time points: one before NCRT and the other after NCRT. The pre-Radiomics study was conducted using pre-therapy MRI images, and the delta-Radiomics study was conducted using pre-therapy and post-therapy MRI images. All rectum MRI examinations were performed using a 3.0-T magnet (Philips Ingenia 3.0T) with a phased array surface coil. Bowel preparation was performed before image acquisition. The following pulse sequences covering the entire tumor were included: (1) axial (perpendicular to the long axis of the rectum) T2-weighted imaging (T2WI). This was obtained with a slice thickness of 3.8 mm, repetition time (TR)/echo time (TE) of 4000 ms/120 ms, a field of view (FOV) of 16 × 16 cm, matrix size of 320 × 256, echo train length (ETL) of 22, and the number of excitation (NEX) of 2.

First, T1 and T2 MRI images were normalized and aligned to facilitate accurate manual segmentation of tumor areas. Then, the regions of interest (ROIs) of the tumors were manually segmented using ITK-SNAP software by two experienced doctors (version 3.8.0; www.itksnap.org). Intraobserver difference of ROI was performed by calculating the Dice ratio. Segmentations with a dice ratio of over 0.90 were considered qualified. For those less than 0.90, the segmentation would be re-evaluated by a third experienced radiologist. An example of tumor segmentation is shown in Figure 2 .

The image processing and radiomics feature extraction were performed using the Pyradiomics tool (version 3.0.1) as previously described (33, 34). A total of 105 three-dimensional Radiomics features from each tumor volume, including 14 shape-based features, 18 first-order features, and 73 texture features, were quantified. The three-dimensional texture features were calculated using gray level co-occurrence matrix (GLCM) (N=22), gray level size zone matrix (GLSZM) (N=14), gray level size zone matrix (GLSZM) (N=16), gray level run length matrix (GLRLM) (N=16), and neighboring gray-tone difference matrix (NGTDM) (N=5).

In the study, two radiomics types were defined: pre-Radiomics and delta-Radiomics. Pre-Radiomics analysis was implemented only using the pre-Radiomics features from pre-therapy MRI. Delta-Radiomics analysis was implemented using the delta-Radiomics features based on changes in radiomic features before and after NCRT, which were post-therapy radiomics features minus the pre-therapy radiomics features. All the imaging features were standardized for the subsequent machine-learning processing.

Then, eXtreme Gradient Boosting (XGBoost) was used to evaluate and select the important features, with the gbtree booster, a max-depth of 10, a lambda of 1, and an eta of 0.01, which was implemented using xgboost Python (version 0.82). Previous studies have also demonstrated that the XGBoost algorithm could be used for processing structured tabular data (35, 36). Based on experience, a feature in more than 10 samples or patients could be more robust for building binary classifiers (37, 38). Therefore, according to the sample size, an appropriate number of features were selected to construct the feature dataset. The sample imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE) (39) to enhance the data and improve the modeling performance. Finally, the pre-Radiomics and delta-Radiomcis models were built using the corresponding features and XGBoost classifier.

Multiple group comparison experiments were performed. First, machine learning models were compared using single-model MRI and multi-modal MRI. The T1, T2, and T1+T2 integrated models were then constructed. Second, original features-based models and resampled features-based models using SMOTE were compared. Then, the pre-Radiomics model and delta-Radiomics model were compared through cross-validation with 5-fold and 10-fold, leave-one-out validation, and independent test. All the models were evaluated using the area under the receiver operator characteristic (ROC) curve (AUC). The degree of importance and statistical differences of valuable features and radiomics prediction scores in post-treatment responses were also assessed.

The relevant statistical analyses and machine learning algorithms were generated using Python (version 3.6.6). Differences between differently distributed variables were compared using T-test or Mann–Whitney U test. XGBoost was performed for feature selection and modeling. The prediction performance of the model was evaluated using the area under the ROC curves and mean AUCs through cross-validation. A Delong test was performed to compare the performance of the models. P-value < 0.05 was considered statistically significant.

Patient demographic characteristics are shown in Table 1 . There was no significant difference in clinical factors between patients with good reactions and poor reactions to LARC. The reliability of results from small sample sizes is usually low (40, 41). Among the 84 study lesions for pre-Radiomics analysis, 28 (33.33%) were classified as having a good reaction, and 56 (66.67%) in the poor reaction group. For delta-Radiomcis analysis with 77 lesions, 27 (35.06%) were good reactions and 50 (64.94%) were poor reactions.

Figures 3A–C , 4A–C show the ROC curves for T1, T2, and T1+T2 models based on pre-Radiomcis and delta-Radiomcis analysis. In pre-Radiomics analysis, the mean AUC of the T1 model was 0.81, that of T2 was 0.73, and that of T1+T2 was 0.89. In delta-Radiomcis analysis, the mean AUC of T1 was 0.77, that of T2 was 0.89, and that of the T1+T2 model was 0.93. Therefore, T1 is more relevant than T2 in pre-Radiomics analysis, but T2 is more relevant in delta-Radiomcis. Based on the Delong test, combining the T1 and T2 models were superior to either model alone (P < 0.05).

Figures 3 , 4 show the ROC curves of the original features-based models and SMOTE-based models. It was found that the oversampled data using SMOTE method is more accurate than original imbalanced data. As shown in Figures 3D–F , 4D–F showed the same diagnostic trends and patterns as before, meaning that the SMOTE technology has superior data mining potential.

Based on the above experiments, we selected T1+T2 as the final radiomics model using SMOTE. The top-5 valuable pre-Radiomics and delta-Radiomics features are shown in Table 2 . All the valuable pre-Radiomics features were from T1 MRI and texture features. T2 MRI had more significance in delta-Radiomics analysis than in pre-Radiomics. Overall, the delta-Radiomics features were more important than pre-Radiomcis features. The types of important pre-therapy imaging features were different from post-therapy features.

The final pre-Radiomics and delta-Radiomics models were built. The 5-fold cross-validation, 10-fold cross-validation, and leave-one-out validation were used to comprehensively evaluate the differences in the prediction performance between the two methods. The prediction performance of the pre-Radiomics and delta-Radiomics models is shown in Table 3 . The predictive accuracy of the delta-Radiomcis model was higher than that of pre-Radiomcis model in 5-fold cross-validation (0.96 vs. 0.93), 10-fold cross-validation (0.95 vs. 0.92), and leave-one-out validation (0.93 vs. 0.90). The accuracy of all radiomics models was higher than 0.90, demonstrating their satisfactory good predictability in predicting the NCRT response of LARC.

A given cutoff prediction value for the models was selected to evaluate their NCRT prediction accuracy. The prediction probability of the machine learning models as the radiomics prediction scores to evaluate the degree of resistance to the treatment response. A higher score means a higher risk of poor reaction. Figure 5 shows the prediction scores of the pre-Radiomics and delta-Radiomics between good and poor response groups in the primary and validation cohort. Both models accurately distinguished responses to NCRT in the primary cohort (P < 0.001) and the validation cohort (P < 0.05).