From: The power of deep learning in simplifying feature selection for hepatocellular carcinoma: a review
Study | Data source | Used algorithms | Input | Output | Key findings |
---|---|---|---|---|---|
Jin, Yao [42]2021 | One center in China | Deep learning-based radiomics | Ultrasound images | Predict HCC development | The DL radiomics model successfully predicted the risk of developing hepatocellular carcinoma (HCC) over five years in the test set, achieving a high AUC of 0.900 |
Brehar, Mitrea [43] | Center in Romania | convolution neural network | Ultrasound images | HCC detection | The CNN model achieved outstanding results in HCC detection, with an AUC of 0.95, accuracy of 91%, 94.4%, and sensitivity of 88.4% |
Shi, Kuang [44] | One center in China | convolution neural network | Computed Tomography images | Focal liver lesion type classification | Applying a convolution neural network to three-phase CT images resulted in a noteworthy performance in differentiating hepatocellular carcinoma from other focal liver lesions (FLLs), with an Area Under the Curve (AUC) of 0.925 |
Wu, White [45] | center in United States | convolution neural network | Magnetic Resonance images | Classification of liver lesions using LI-RADS | The convolution neural network model delivered exceptional results in LI-RADS grading of liver tumors, exhibiting an AUC of 0.95. With an accuracy of 90%, the model demonstrated a high sensitivity of 100% and a positive predictive value (PPV) of 83.5% |
Zhen, Cheng [46] | center in China | convolution neural network | Magnetic Resonance images | Liver tumor type classification | The integration of clinical data with a convolution neural network resulted in a highly accurate classification of hepatocellular carcinoma, with an AUC of 0.985 and a strong agreement rate of 91.9% compared to pathology |
Wang, Jian [47] | center in China | convolution neural network | Magnetic Resonance images | Microvascular invasion in hepatocellular carcinoma | The combination of deep features from MRI images achieved an AUC of 0.79 when predicting macrovascular invasion (MVI) in HCC patients |
Jiang, Cao [48] | center in China | convolution neural network | Computed Tomography images | Microvascular invasion in hepatocellular carcinoma | convolution neural network achieved an AUC of 0.906 for the prediction of MVI. Mean survival was better in the group without MVI |
An, Jiang [49] | center in China | convolution neural network | Magnetic Resonance images | detects the amount of normal tissue that is destroyed around a tumor during cancer ablation therapy (ablative margin) | The deep learning model provided accurate estimations of ablative margins and effectively assessed the risk of tumor recurrence at the original site |
Liu, Xu [50] | center in China | convolution neural network | Computed Tomography images | Survival rate after transarterial chemoembolization (TACE) | A higher DL score served as an independent prognostic factor, accurately predicting overall survival with AUC values ranging from 0.85 to 0.90 |
Liu, Liu [51] | center in China | convolution neural network | Ultrasound images | The outcome of Transarterial Chemoembolisation | The model successfully predicted the tumor Outcome of transarterial chemoembolization (TACE) with a remarkable area under the curve of 0.93 |
Peng, Kang [52] | three centers in China | convolution neural network | Computed Tomography images | The outcome of Transarterial Chemoembolisation | In two separate validation cohorts, the deep learning model exhibited accuracies of 85.1% and 82.8% when predicting the outcome of transarterial chemoembolization (TACE) |
Zhang, Xia [53] | three centers in China | convolution neural network | Computed Tomography images | Overall survival of patients treated with transarterial chemoembolization (TACE) and sorafenib | The deep learning signature demonstrated a C-index of 0.714 in accurately predicting overall survival in hepatocellular carcinoma patients who underwent treatment with TACE and sorafenib |
Mitrea, Brehar [54] | GE7 dataset and GE9 dataset | fusion between the convolution neural network and CML methods | ultrasound image | enhancing the HCC automatic recognition performance | The integration of CNN-based techniques with traditional machine learning methods, leveraging advanced texture analysis, has demonstrated remarkable effectiveness, yielding classification accuracies surpassing 95% in numerous scenarios |
Lai, Wu [55] | China Medical University Hospital | ResNet-18 convolutional neural network | Computed Tomography images and FDG PET-CT images | Overall Survival Prediction in Patients with Hepatocellular Carcinoma Based on 18F-FDG PET-CT Images | The developed prognostic model combined FDG PET-CT images and FDG CT images, Leading to better performance than using CT images alone (0.807 AUC vs. 0.743 AUC). The model utilizing FDG PET-CT images exhibited slightly higher sensitivity than the model relying solely on CT images (0.571 SEN vs. 0.432 SEN) |
Sun, Shi [56] | Hospital of Harbin Medical University (Harbin, China) | (LASSO) regression for feature selection Deep learning and radiomic | CECT images | Predicting Treatment Response to Transarterial Chemoembolization (TACE) in Patients with Hepatocellular Carcinoma | The DLRC model was created by incorporating 19 quantitative radiomic features, 10 deep learning features, and 3 clinical factors. In the training cohort, the DLRC model achieved an AUC of 0.937 (95% confidence interval [CI], 0.912–0.962), while in the validation cohort, it achieved an AUC of 0.909 (95% CI, 0.850–0.968) |