Introduction：Colorectal cancer (CRC) is a prevalent malignancy within the digestive system, exhibiting a substantial global incidence[1].Neoadjuvant chemoradiotherapy (nCRT) plays a pivotal role in the treatment of locally advanced rectal cancer (LARC) by enhancing patient survival and facilitating tumor reduction[2]. However, not all patients derive benefits from nCRT[3,4]. Therefore, accurate prediction of treatment efficacy is crucial for rectal cancer patients in guiding treatment decisions and predicting prognosis.Increasing evidence suggests that texture features in MRI images provide qualitative and quantitative information about rectal cancer tissue. However, manual determination of these features is labor-intensive and prone to error, which limits its clinical applicability[5].With the advancement of computer artificial intelligence technology, machine learning (ML) has gained traction in predicting the efficacy of neoadjuvant therapy for rectal cancer, significantly enhancing accuracy and efficiency[6]. Nonetheless, the challenge of selecting appropriate features persists. The aim of this study is to predict the efficacy of nCRT for rectal cancer based on wavelet texture feature analysis of magnetic resonance images. Wavelet transformation can effectively extract texture information from images, assessing tumor tissue complexity and variation by analyzing features such as maximum probability, minimum value, kurtosis, and maximum correlation coefficient.
Methods：A retrospective collection was conducted involving 62 patients with locally advanced rectal cancer between June 2016 and December 2020. Clinical characteristics of the patients were collected. Efficacy assessment was conducted according to the revised RECIST (1.1 version) criteria.MRI examinations were performed on a 3.0-T system (Signa HDXT, General Electric Healthcare, Waukesha, WI, USA). For the T1WI sequence, the following parameters were used: repetition time (TR) of 500 ms, echo time (TE) of 7.2 ms, number of excitations (NEX) of 1, field of view (FOV) of 32 cm × 32 cm, slice thickness of 6 mm, and no interslice gap. The T2WI sequence was acquired with parameters including a TR of 3500 ms, TE of 109.1 ms, NEX of 4, FOV of 24 cm × 24 cm, slice thickness of 3 mm, and no interslice gap. IVIM-DWI was performed using 10 different b-values (0, 10, 20, 50, 100, 200, 400, 800, 1200, 2000 s/mm²), with a TR of 4000 ms, TE of 65 ms, and NEX of 6. Wavelet texture feature parameters of all lesions in patients' MRI images were extracted. The Random Forest Classifier model was employed for feature selection, followed by classification learning using the XGBoost classifier.
Results ：A total of 62 patients with 1656 magnetic resonance wavelet texture feature parameters from all lesions were obtained. Feature selection using the Random Forest Classifier identified the most important four wavelet texture features, which included DWI wavelet-HHH glcm Maximum Probability, DWI wavelet-LLL firstorder Minimum, T2 wavelet-LHH firstorder Kurtosis, and T2 wavelet-HLL glcm MCC. Utilizing the XGBoost classifier for predictive classification yielded a final accuracy of 84.21% and an AUC value of 0.90. The precision, recall, and F1 score for the responsive treatment group were 0.86, 0.92, and 0.89, respectively. For the non-responsive treatment group, the precision, recall, and F1 score were 0.80, 0.67, and 0.73, respectively. The macro-averaged values of precision, recall, and F1 score across all labels were 0.83, 0.79, and 0.81, respectively.
Discussion：Machine learning can automatically extract key features from extensive data, construct prediction models, provide fast and accurate prediction results, alleviate clinicians' burdens, and offer more reliable references for clinical decision-making.In this study, a Random Forest model was employed to extract key features from a vast array of wavelet texture features. The Random Forest model is capable of handling high-dimensional datasets and a multitude of features, exhibiting excellent scalability[7]. It is adept at managing non-linear relationships and interaction effects. It performs well when dealing with mixed data containing discrete and continuous features, aiding in the identification of features of higher importance to the target variable.Random Forest can estimate the importance of features, measuring the contribution of features to the model. By selecting and analyzing texture features from different regions of MRI images, we achieved a predictive accuracy of 84.21% and an AUC value of 0.90 for nCRT efficacy prediction of localized advanced rectal cancer in this experiment. Therefore, this study demonstrates that on our dataset, the XGB Classifier model exhibits excellent predictive performance.
Conclusion：The model established based on wavelet texture feature analysis of magnetic resonance images can effectively predict the efficacy of neoadjuvant radiochemotherapy for rectal cancer patients. The XGB Classifier model demonstrates excellent performance in handling small, imbalanced datasets, thus showing promising clinical application value.

Acknowledgements

No acknowledgement found.