RADIOMICS | Post

Deep Learning: A multi-center and prospective validation study

Please click here to download the model.

MVI-Net: This study proposed a combined deep learning model (CDLM) for microvascular invasion (MVI) prediction in hepatocellular carcinoma. This model is designed based on the ResNet18 and utilized transfer learning to initialize the first 14 convolutional layers of the model. We trained the CDLM^EOB-MRI model in 987 MR images from 329 hepatocellular carcinoma patients in two hospital, and validated its performance in another independent hospital including 166 patients. We trained the CDLM^CE-CT model in 918 CT images from 306 hepatocellular carcinoma patients in two hospital, and validated its performance in another independent hospital including 115 patients.

The CDLM^EOB-MRI model achieved encouraging predictive performance in both the primary cohort (AUC = 0.962) and the independent validation cohort (AUC = 0.842). The CDLM^CE-CT model achieved good predictive performance in both the primary cohort (AUC = 0.842) and the independent validation cohort (AUC = 0.736).

To use this model, users only need to feed the region of interest (ROI) of the tumor MR or CT image and corresponding clinical factors into the model. The CDLM will give the MVI probability for the patient and highlight the suspicious tumor region that are most related to MVI status, which is very easy to use in application.

Email: gudongsheng2016@ia.ac.cn, weijingwei2014@ia.ac.cn

Figure 1. The deep learning model for a single sequence.

Figure 2. The combined deep learning model (CDLM).

Figure 3. Predictive performance of the developed models. a-d) ROC curves of the developed models in the EOB-MRI training cohort, CE-CT training cohort, validation cohort 1 with 166 patients, and validation cohort 2 with 115 patients, respectively. e-f) Decision curves of the CDLM^EOB-MRIand CDLM^CE-CT, respectively. The red, green, and blue lines show the benefits of using the developed CDLM, radiological model and clinical model, respectively. The gray and black line represent the benefit of treating all patients as MVI-positive and negative type, respectively. g-h) Violin graphs of the CDLM^EOB-MRI and CDLM^CE-CT, respectively.

Figure 4. The convolutional filters (features) of the deep learning model.

Deep learning based radiomics for preoperative microvascular invasion and high-risk area detection with comparison between gadoxetic acid-enhanced MR and contrast-enhanced CT: a multi-center and prospective validation study

Deep Learning: A multi-center and prospective validation study