This is the first study to use radiomics from CT images of UC patients at admission to predict indications for surgery. Combining radiomics data with clinical information at admission yielded a high predictive power of approximately 90% (AUC 0.896 (95% CI 0.841–0.951) in the training cohort). This predictive model was also highly robust in the validation cohort, with an AUC of 0.868 (95% CI 0.729-1.000). This study suggests that radiomics may be able to predict discharge outcomes for UC patients from CT at admission.
Although CT scans are frequently performed in the practical treatment of UC patients, their usefulness is not clear. The percentage of UC patients who undergo a CT scan at the time of their outpatient visit has been reported to range from 8–89% [25–28]. The more recent the report, the higher the percentage of CT performed. The probability of significant findings on CT in symptomatic UC patients is low, and some are skeptical about performing CT routinely in terms of radiation exposure[23–26]. On the other hand, some reports indicate that mural stratification on CT at admission can be used to predict surgical intervention [22]. In recent years, CT entrography findings have been shown to correlate well with the severity of illness and pathology [25], but it is difficult to perform enterography at the time of emergency hospitalization due to pretreatment. In this study, the use of radiomics to extract high-dimensional information from noncontrast CT images improved the accuracy of predicting indications for surgery, further suggesting the usefulness of CT examinations. Noncontrast CT scans can be performed in a short time, making them easier to perform in emergency situations, and it is expected that deepening this model will have an important role in determining treatment strategies for severe UC.
The development of radiomics-based disease classification and prognostic models has been attempted for various diseases [16–19, 29–31]. Radiomics has the advantage of being able to express image characteristics numerically, and it is also expected to be useful in quantitatively evaluating qualitative decisions that have been made by professionals. It has also been shown to contain microscopic information such as pathological findings and genetic information and is expected to be useful as a digital biopsy [13, 14]. Previous IBD studies have suggested that radiomics can differentiate CD from UC, predict disease stage, and predict treatment response [20, 21, 32–37]. Radiomics alone provides good discrimination but combining it with clinical data further improves discrimination[20, 21, 36, 37]. Studies have also been conducted using other modality imaging, such as MRI and CT enterography[32–37]. There are many issues to be addressed in studies using radiomics, such as appropriate imaging modalities, image preprocessing methods, and ROI definition, and further validation is desirable. Radiomics data are often analyzed in combination with machine learning because of the enormous amount of information represented. Image recognition and analysis using deep learning has also been rapidly put to practical use in recent years. From a study comparing radiomics analysis and deep learning for intestinal fibrosis classification in CD, scholars reported that both approaches exhibited a high degree of performance with no difference in discriminative ability[32]. Radiomics has the advantage of allowing humans to define regions of interest and extract specific data. Deep learning, on the other hand, does not require annotation and is a simpler process. Establishment of optimal image analysis methods in IBD will enable more accurate medical care.
Early determination of the indication for surgery in UC is necessary to improve the safety of surgery. In general, a good preoperative general condition is important for reducing the incidence of perioperative complications. In the preoperative management of UC, improvement of nutritional status, reduction in steroid use, and prevention of thrombosis are recommended. Prediction of the indication for surgery is an important issue because if surgery is not predictable and medical therapy is continued, surgery may have to be performed when the patient's general condition is at its worst. Currently, the possibility of surgery is predicted sensibly by clinicians based on blood sampling data at admission, the severity of UC, and the history of drug treatment prior to admission. In this study, the regression model combined with the clinical findings also showed good discriminative power. Radiomics provided a complementary dataset that had the effect of increasing the accuracy of classification by clinical findings.
This study is affected by several limitations. First, ROI selection was performed manually only. There are no established methods for automatic or semiautomatic segmentation of nonpretreated intestinal tracts, and the development of these methods could lead to more accurate predictive models. Second, this was a retrospective cohort study, and there were many selection biases. In general, patients who undergo CT are relatively severely ill, and the selection of patients depended on whether CT imaging was performed. However, given considerations of radiation exposure, it is ethically difficult to perform CT on all hospitalized UC patients; prospective studies should be conducted to establish certain criteria for CT imaging.