Introduction
Hepatocellular carcinoma (HCC) ranks second as the major cause of cancer-related deaths globally and is the sixth most common cancer in the world. Its incidence has continuously increased in recent years, and approximately 850,200 new cases of HCC are annually diagnosed worldwide [
1,
2]. Less than 30% of patients with HCC are eligible for potentially curative therapies, such as transplantation, resection, or ablation [
3,
4]. For selected patients who are not suitable for such interventions, but have liver-confined disease, preserved liver function, and good performance status, transarterial chemoembolization (TACE) is recommended according to international guidelines [
5‐
9].
Although repeated TACE procedures are often needed, the initial response effectively predicts the overall survival (OS) because the best response cannot always be achieved after one session of TACE, especially in large tumors. Moreover, the achievement of a treatment response at an early time point is the robust predictor for favorable outcomes [
10]. The texture analysis based on contrast-enhanced magnetic resonance imaging (MRI) before TACE may act as imaging biomarkers to predict an early response from patients with HCC. The highest accuracy for complete response (CR) group and the non-complete response (NCR) was 0.76 [
11]. A pretherapeutic dynamic CT texture analysis can also be valuable in predicting the therapy response of HCC to TACE. Higher arterial enhancement and GLCM (gray-level co-occurrence matrix) moments, lower homogeneity, and smaller tumor size are significant predictors of complete response (CR) after TACE [
12]. However, based on the traditional statistics and machine learning, the accuracy of this method is limited. On the condition of optimal cutoff values for predicting a CR to TACE in the receiver operating characteristic (ROC) curves, the highest AUC of texture parameters was 0.72. Furthermore, most studies focused on the two classifications (CR or NCR) and the prediction of four classifications (CR, PR, SD, and PD) using CT images is unclear. Therefore, a more effective model to accurately identify patients who would have an initial response after TACE therapy is urgently needed to facilitate individualized treatment strategies.
Deep learning has recently gained attention as a technique for realizing Artificial intelligence (AI) [
13‐
15]. Several types of deep-stacked artificial neural networks, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been proposed and judiciously used in various fields. Deep CNNs are especially recognized as demonstrating high performance for image recognition tasks [
16,
17]. Some initial successes in applying deep learning to the assessment of radiological images have been witnessed [
18‐
21]. A study used a deep learning algorithm to non-invasively predict the IDH (Isocitrate Dehydrogenase) status within a multi-institutional dataset of low- and high-grade gliomas [
22]. Deep learning also shows the potential to stage liver fibrosis based on radiological images [
23]. However, there is embarrassed that performing deep learning often faces a shortage of medical data, especially in radiological images of patients undergoing treatment. Transfer learning, which is a feasible deep learning technique for addressing a lack of image data, has been proven a highly effective technique, particularly in the case of limited medical images [
24,
25]. The models have been used to distinguish the features of the medical images in a much faster manner and with significantly fewer training medical images [
13].
In this study, based on the CT images from three independent centers, we aimed to investigate a deep learning algorithm to precisely and non-invasively evaluate the different therapy response in HCC patients before the TACE treatment.
Discussion
In this study, we demonstrated a novel application of deep learning to predict the response of TACE therapy in a three-institution dataset of HCC. As far as we know, this is the first time that the deep learning model based on radiological images is used to predict the four responses of interventional treatment in liver cancer. This algorithm may facilitate deep learning techniques for the medical field of precise therapy oncology. Based on the pretreatment ROI images of patients with HCC, this utility model of deep learning is a potential method for predicting the response of TACE therapy.
According to the BCLC stage system, TACE is recommended for patients with HCC with BCLC stage B [
6,
30‐
32]. In patients with stage C, TACE therapy is also a frequent and important application treatment, especially in comprehensive treatment [
33‐
35]. Recent studies revealed that the therapy response at first chemoembolization is a good predictor for the favorable outcome in hepatocellular carcinoma [
10,
36]. However, no report was associated with the prediction response of TACE therapy in the field of hepatocellular carcinoma via deep learning of CT images. Previous studies showed that different clinical risk factors (e.g., tumor size) perform well in prediction [
37‐
39], but the precise estimation of four therapy responses remains challenging in clinical settings and difficult to implement. A new method involving a radiomics approach based on radiological images (e.g., CT, MR, and PET-CT) is also currently being applied in various tumors. It extracts radiographic features from conventional images and includes the features of tumor shape, texture, intensity, and wavelet transform characteristics [
40‐
46]. However, numerous pre-engineered features are artificial design features. This may lead to poor reproducibility and nonredundant radiomics features (RFs) for CT images because of the variable scan parameters of different types of imaging equipment [
47]. The application of radiomics also relies on traditional machine learning techniques. Unlike the above method, the algorithm of deep learning can directly learn predictive features from the images and potentially greatly increase the robust accuracy in these radiological images [
48,
49].
Previous studies mostly focused on deep learning based on the small segmentation patches of each ROI image to enhance the size of the sample and frequently showed a significant AUC [
50‐
52]. However, the scenario of predicting the labels of entire ROI images was often ignored. Therefore, we used an algorithm of transfer learning to evaluate the treatment response via the whole CT-ROI patches. Comparing with previous study, high AUCs of predicting CR, PR, SD, and PD in the therapy response of TACE were observed among the three cohorts [
12]. This result indicated that our transfer learning model performed well in predicting different therapy responses using CT images from three independent centers. The result of the confusion matrix presented significantly high accuracies of prediction in the NFH, ZHHAJU, and SYUCC cohorts and was distinct from previous report [
11]. Interestingly, we found the accuracy for the training cohort was lower than for the validation cohort 1 (84.3% vs. 85.1%). We speculated the phenomenon was correlated with a small sample size of patients in validation cohort 1. Increasing number of patients would potentially reduce the validation accuracy. Misclassified CR patches by the deep learning model were more observed in PR patches than in SD and PD patches in the training cohort (1.5%) and validation cohorts 1 (1.7%) and 2 (1.1%). Meanwhile, misclassified PD patches were more frequently found in PR patches than in SD and CR patches. The precision probability of preoperatively predicting the four therapy responses (i.e., CR, PR, SD, and PD) via each ROI patch was calculated and found useful in individualized clinical treatment. We further investigate the prediction of objective response (response or non- response) in patches or patients and also found high accuracies in the three cohorts. This finding demonstrated that the deep learning model based on CT images may help doctors recognize patients who would acquire well or poor initial response of TACE therapy.
However, our study has several limitations. First, the sample size of patients with HCC was relatively small, and this was a retrospective research. A much larger database of the prospective study would be collected from more centers in the future. Second, we trained and validated all the patches of the 2D CT images from three medical centers. Because of 3D patches’ potential of having more context information, we speculated the 3D CT patches had an accuracy higher and a better model quality than that of the 2D CT patches. The 3D CT patches would be investigated in the next step. Third, the correlation between the biological processes (e.g., differential gene expression and pathway) and the prediction results of deep learning networks in HCC was unknown and should be analyzed in the future. Fourth, the ROIs were drawn manually in our study. Lesions selected from different abdominal radiologists might have various differences, impacting on disease classification. We would use the combination of the algorithm for HCC segmentation and ResNet50 model to automatically predict the outcome of TACE therapy in the following study.
In summary, the deep learning model based on CT images would potentially serve as a new tool for predicting the therapy response of patients undergoing TACE treatment. Our method using transfer learning for predictive classification of radiological images may also be used to determine more precise clinical treatments in other malignant tumors.
Acknowledgments
(I) Conception and design: Li Liu and Jie Peng; (II) administrative support: none; (III) provision of study materials or patients: Li Liu and Jie Peng; (IV) collection and assembly of data: Jie Peng, Shuai Kang, Hangxia Deng, Yikai Xu, Jing Zhang, Wuxing Gong, and Jingxian Shen; (V) data analysis and interpretation: Li Liu, Jie Peng, Zhengyuan Ning, and Jinhua Huang; (VI) manuscript writing: all authors; and (VII) final approval of manuscript: all authors.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.