We applied the DL model to predict postoperative early recurrence of iCCA. We have successfully demonstrated high performance in the prediction of postoperative recurrence using plain preoperative CT images. The accuracy of the DL model far exceeded that of the binary logistic regression analysis (AUC, 0.994 vs. 0.770). This report represents the first study in which a DL model based on CT images is used to predict postoperative recurrence in iCCA. Our results may yield a novel insight into personalized treatment strategies, including neoadjuvant and adjuvant chemotherapy, in iCCA management.
Adjuvant chemotherapy is certainly expected to increase the survivorship of patients with iCCA 19–21. Isolating results has been a challenge as past prospective randomized trials have included not only iCCA but also other bile duct cancers 21–23. Furthermore, the indication for adjuvant chemotherapy in those studies was heterogeneity. In short, the selection criteria of susceptible individuals for adjuvant chemotherapy is not well established. To address this issue, Jeong et al. showed the usefulness of an AI framework in the prognostic estimation and stratification of susceptible individuals for adjuvant treatment after resection in iCCA patients. In contrast, we intended to predict postoperative recurrence directly. Our model, which can directly predict early recurrence, would be used to predict who should receive adjuvant chemotherapy based on their risk of recurrence.
Liang et al. conducted a single-center retrospective study and built a radiomics nomogram to predict early recurrence of iCCA after surgical resection 24. Their nomogram, using preoperative arterial-phase contrast-enhanced magnetic resonance imaging (MRI), achieved an AUC of 0.82 and 0.77 in the training and validation cohorts, respectively. They used manual engineered features and selected the earlies recurrence-related features using a least absolute shrinkage and selection operator logistic regression analysis. Zhao et al. used radiomics from MRI to predict early recurrence. Their radiomics model showed a preferable predictive performance (AUC 0.889) 25. Compared with the previous radiomics model using MRI, our model, which is based on DL features, achieved higher predictive performance (AUC 0.994).
Based on our results, which perform in such a highly predictive manner with the model addressing postoperative early recurrence, we propose a new concept in iCCA management. Though we need to discuss further which population, patients, those with or without early recurrence, is fit for adjuvant chemotherapy, achieving quite high levels of predictive accuracy, compared to conventional methods, can provide valuable information for determining adjuvant therapy and developing surgical plans, thereby facilitating pretreatment decisions.
Thanks to the advantages reaped from DL, we physicians, can easily apply computer-aided diagnosis 16,26. Deep learning algorithms, such as CNN, have been widely used in the field of image diagnosis and prediction owing to their being fast, accurate, and reproducible 26,27. CNN can uncover details in medical images that human experts cannot find, and automatically render a quantitative assessment 28. Generally, even expert radiologists and surgeons cannot always access meaningful findings that would enable physicians to decide on a treatment strategy from plain CT images. In fact, there have been no reports or guidelines that recommend using plain CT images for risk assessment of postoperative recurrence in iCCA. Several lines of evidence, including our study, can lead to a paradigm shift in the recognition of AI in the field of iCCA treatment.
The present study has several limitations. This is a retrospective study. In addition, although this is a pilot study, the patient population was small. However, our model achieved high predictive performance. If we had access to additional training data from a large cohort, we could achieve even higher prediction accuracy and generality. To establish clinical applications, sufficient datasets are fundamental requirements. A novel AI approach based on analyzing a huge database, such as national or regional datasets, would be attractive to both clinicians treating iCCA and their patients. An accurate and robust prediction model can ultimately contribute to a better prognosis in iCCA patients.
There was the question of possible lack of homogeneity in CT techniques over the past 20 years that has been a point of contention (Supplemental Table 1). Nevertheless, our model achieved high predictive performance. These results suggested that relative heterogeneity of CT techniques may not be a big issue because of the handling of huge information from CT images through DL. Certainly, homogeneity of CT techniques would be preferable. However, it would not be practical in a real clinical setting for all patients to undergo CT exams using the same scanner and technique. In short, the use of a diverse set of CT acquisitions was not a limitation, it was a benefit to the study.
In conclusion, our DL model, using plain preoperative CT images of iCCA, exhibited high predictive performance in projecting postoperative early recurrence. The present multicenter study has provided a novel approach to predict recurrence after surgery. This model may help clinicians in the selection of patients for neoadjuvant and/or adjuvant therapy. This approach can contribute to personalized strategies in iCCA treatment. To establish a clinical application, conducting a study using a huge dataset, such as national dataset, is the hope for the future.