This study aimed to establish a variety of radiomics models to predict the treatment effect of chemotherapy for STS lung metastases. A total of 6 kinds of radiomics models were established, of which the model with the best performance in the training and test groups was the decision tree classifier. The AUCs of the models were all greater than 80%. Except for the poor specificity of 55.6% in the test group, all the other indicators showed good predictive value of the model. Radiomics has been widely used in the evaluation of the treatment efficacy in various tumours. It can use advanced quantitative feature analysis methods to extract a large amount of data from medical images for digital analysis and obtain a large amount of high-fidelity target information. Multiple layers of the tumour are comprehensively evaluated to reflect the changes before and after tumour treatment . In previous studies, radiomics has been widely used in the evaluation of the efficacy of neoadjuvant chemoradiotherapy for STS. Gao et al  and Blackledge et al  used a small sample of conventional radiomics to predict the curative effect of neoadjuvant therapy for STS and achieved certain research results. Crombe first proposed the application of delta -radiomics to predict the efficacy of neoadjuvant therapy for STS. He analysed the value of delta-radiomics based on T2-weighted sequences in predicting pCR in STS patients before and after neoadjuvant therapy in 65 patients . In a recent study, Peeken et al  retrospectively studied 156 patients treated with neoadjuvant chemoradiotherapy and established a delta-radiomics model to predict the efficacy of neoadjuvant therapy for STS, which further confirmed the advantages of radiomics in evaluating the treatment efficacy for STS. However, in previous studies, radiomics was mostly used to evaluate the efficacy in primary lesions of specific diseases, and the evaluation of radiomics in the lung was mostly based on lung cancer. No evidence was found for either STS or other primary tumours. The literature describes radiomics research on the efficacy evaluation of intrapulmonary metastases, and this study is pioneering and advanced .
In this study, the RECIST 1.1 criteria were selected as the control criteria because they are currently the most widely used clinical efficacy evaluation criteria for solid tumours . Although the RECIST 1.1 criteria are controversial in the evaluation of neoadjuvant therapy for STS, some STSs that respond biologically to radiotherapy may not shrink due to tumour enlargement due to necrosis, intratumoural haemorrhage, and cystic degeneration , but this study is a study of lung metastases, and the above situation does not apply. In previous studies, in the radiomics study of multiple lesions, a single lesion was usually selected for study in a single patient . However, the delineation of ROIs in this study was different because there may be a single lung metastasis or multiple lung metastases in patients with STS. Therefore, the requirements of RECIST1.1 were used in the delineation of ROIs, and a target organ could be delineated at most two times. For each target lesion, single or multiple lesions can be selected for delineation, and the corresponding curative effect results were the same, which not only increases the sample size of the ROI but also makes the research more in line with clinical reality. The two radiomics features finally screened in this study were wavelet-HHH_First Order Mean and wavelet-LHL_GLRLM Long Run Low Grey Level Emphasis, which were both radiomics features obtained from wavelet transform. The mean is expressed as the average value of the signal intensity of all pixels in the ROI, which can reflect the regularity of the image texture. The greyscale run length matrix (GLRLM) is defined as the length of the number of pixels, i.e., consecutive pixels with the same greyscale value, where LRLGLRE means measuring the joint distribution of long run lengths with lower greyscale values, short-run dominance versus long-run dominance. The advantage reflects the smoothness and roughness of the image. The greater the advantage of the short run is, the smoother the texture of the image, and the greater the advantage of the long run is, the rougher the texture of the image .
From the perspective of clinical problems, this study selected the lung metastasis of STS as the research object because the probability of long-term lung metastasis of STS is very high, and chemotherapy is the key strategy for palliative treatment after lung metastasis of STS. Some studies reported that doxorubicin combined with ifosfamide fails to show a very obvious overall survival benefit for the treatment of metastatic STS . However, as a chemotherapy regimen used worldwide for nearly 40 years, doxorubicin combined with ifosfamide remains the standard first-line treatment for patients with pulmonary metastases from STS. The epirubicin used in this study is an isomer of doxorubicin, which has similar efficacy value and less cardiotoxicity. There are many factors that may affect the lung metastasis of STS, such as the grade of STS (according to the French sarcoma organization FNCLCC, grades II and III are considered to have more metastatic potential), the size and depth of the primary tumour, etc. . These factors may also be clinical factors that affect the efficacy and response of STS after chemotherapy. Some previous studies have used radiomics scores and clinical data to establish a nomogram to predict the efficacy and long-term survival of STS after treatment . In this study, many patients had experienced metastasis when they were admitted to the hospital, and the clinical data in other hospitals were incomplete or unverified. Therefore, a prediction model combining radiomics and clinical data could not be established. Here, we focused on the efficacy assessment and prediction of lung metastases. In this study, some patients did not choose epirubicin combined with ifosfamide as the first-line treatment regimen, and other treatment regimens were unevenly used, making it difficult to perform good stratification. To control the influencing factors of different drugs, drug selection was limited to epirubicin combined with ifosfamide. In future research, we can try to compare the efficacy of multiple drug efficacy evaluation models. Accurate clinical data and good stratification as well as clinical imaging combined with models to predict efficacy are our focus and future research direction.
There are certain limitations in this study. First, there were few enrolled patients, and there was controversy in the statistical analysis of the clinical data. Second, the sample data were from a single-centre retrospective study. ROI segmentation was performed by manual or semiautomatic delineation, some layers of tumour lesions may be poorly demarcated from normal tissue, and there may be potential errors that may cause a certain deviation. Moreover, the clinical information included in the study was limited, and the possible influencing factors of various clinical aspects on the curative effect cannot be carefully incorporated. Finally, due to the variety of follow-up scanning methods for the enrolled patients and the uneven number of images of the patients, they could not be reasonably grouped, so this study only selected images of enhanced scans for research. In future studies, we will conduct a prospective, multicentre, large-sample imaging study and include a complete and refined analysis of clinical factors.
In conclusion, CT radiomics has certain value for noninvasively predicting the therapeutic efficacy of epirubicin combined with ifosfamide in patients with STS lung metastases, and our findings suggest that CT radiomics with multiple ROIs can be used to study multiple lung metastases. This study can help guide individualized treatment strategies for pulmonary metastases from STS. Although validation in a large-scale prospective study is needed, our preliminary findings suggest that CT radiomics has the potential to be used as a noninvasive biomarker to predict the efficacy of lung metastases from STS.