Artificial intelligence (AI) models for the ultrasonographic diagnosis of liver tumors and comparison of diagnostic accuracies between AI and human experts

Diagnostic Performance of Artificial Intelligence in Detection of Hepatocellular Carcinoma: A Meta-analysis

Due to the increasing interest in the use of artificial intelligence (AI) algorithms in hepatocellular carcinoma detection, we performed a systematic review and meta-analysis to pool the data on diagnostic performance metrics of AI and to compare them with clinicians' performance. A search in PubMed and Scopus was performed in January 2024 to find studies that evaluated and/or validated an AI algorithm for the detection of HCC. We performed a meta-analysis to pool the data on the metrics of diagnostic performance. Subgroup analysis based on the modality of imaging and meta-regression based on multiple parameters were performed to find potential sources of heterogeneity. The risk of bias was assessed using Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) and Prediction Model Study Risk of Bias Assessment Tool (PROBAST) reporting guidelines. Out of 3177 studies screened, 44 eligible studies were included. The pooled sensitivity and specificity for internally validated AI algorithms were 84% (95% CI: 81,87) and 92% (95% CI: 90,94), respectively. Externally validated AI algorithms had a pooled sensitivity of 85% (95% CI: 78,89) and specificity of 84% (95% CI: 72,91). When clinicians were internally validated, their pooled sensitivity was 70% (95% CI: 60,78), while their pooled specificity was 85% (95% CI: 77,90). This study implies that AI can perform as a diagnostic supplement for clinicians and radiologists by screening images and highlighting regions of interest, thus improving workflow.

Integrating Artificial Intelligence Into the Visualization and Modeling of Three-Dimensional Anatomy in Pediatric Surgical Patients

BACKGROUND

Pediatric surgeons often treat patients with complex anatomical considerations due to congenital anomalies or distortion of normal structures by solid organ tumors. There are multiple applications for three-dimensional visualization of these structures based on cross-sectional imaging. Recently, advances in artificial intelligence (AI) applications and graphics hardware have made rapid 3D modelling of individual structures within the body accessible to surgeons without sophisticated and expensive hardware. In this report, we provide an overview of these applications and their uses in preoperative planning for pediatric surgeons.

METHODS

Deidentified DICOM files containing cross-sectional imaging of preoperative pediatric surgery patients were loaded from an institutional PACS database onto a secure PC with dedicated graphics and AI hardware (NVIDIA Geforce RTX 4070 laptop GPU). Visualization was obtained using an open-source imaging platform (3D Slicer). AI extensions to the platform were utilized to delineate the anatomy of interest.

RESULTS

Segmentations of skeletal and visceral structures within a scan were obtained using the TotalSegmentator extension with an average processing time under 5 min. Additional AI modules were utilized for providing detailed mapping of the airways (AirwaySegmentation), lungs (Chest Imaging Platform), liver (SlicerLiver), or vasculature (SlicerVMTK). Other extensions were used for delineation of tumors within the hepatic parenchyma (MONAI Auto3DSeg) and hepatic vessels (RVesselX).

CONCLUSION

AI algorithms for image interpretation and processors dedicated to AI functions have significantly decreased the technical and financial requirements for obtaining detailed three-dimensional images of patient anatomy. Models obtained using AI algorithms have potential applications in preoperative planning, surgical simulation, patient education, and training.

LEVEL OF EVIDENCE

V, Case Series, Description of Technique.

Artificial intelligence in hepatocellular carcinoma diagnosis: a comprehensive review of current literature

BACKGROUND AND AIM

Hepatocellular carcinoma (HCC) diagnosis mainly relies on its pathognomonic radiological profile, obviating the need for biopsy. The project of incorporating artificial intelligence (AI) techniques in HCC aims to improve the performance of image recognition. Herein, we thoroughly analyze and evaluate proposed AI models in the field of HCC diagnosis.

METHODS

A comprehensive review of the literature was performed utilizing MEDLINE/PubMed and Web of Science databases with the end of search date being the 30th of September 2023. The MESH terms "Artificial Intelligence," "Liver Cancer," "Hepatocellular Carcinoma," "Machine Learning," and "Deep Learning" were searched in the title and/or abstract. All references of the obtained articles were also evaluated for any additional information.

RESULTS

Our search resulted in 183 studies meeting our inclusion criteria. Across all diagnostic modalities, reported area under the curve (AUC) of most developed models surpassed 0.900. A B-mode US and a contrast-enhanced US model achieved AUCs of 0.947 and 0.957, respectively. Regarding the more challenging task of HCC diagnosis, a 2021 deep learning model, trained with CT scans, classified hepatic malignant lesions with an AUC of 0.986. Finally, a MRI machine learning model developed in 2021 displayed an AUC of 0.975 when differentiating small HCCs from benign lesions, while another MRI-based model achieved HCC diagnosis with an AUC of 0.970.

CONCLUSIONS

AI tools may lead to significant improvement in diagnostic management of HCC. Many models fared better or comparable to experienced radiologists while proving capable of elevating radiologists' accuracy, demonstrating promising results for AI implementation in HCC-related diagnostic tasks.

Artifact reduction in photoacoustic images by generating virtual dense array sensor from hemispheric sparse array sensor using deep learning

PURPOSE

Vascular distribution is important information for diagnosing diseases and supporting surgery. Photoacoustic imaging is a technology that can image blood vessels noninvasively and with high resolution. In photoacoustic imaging, a hemispherical array sensor is especially suitable for measuring blood vessels running in various directions. However, as a hemispherical array sensor, a sparse array sensor is often used due to technical and cost issues, which causes artifacts in photoacoustic images. Therefore, in this study, we reduce these artifacts using deep learning technology to generate signals of virtual dense array sensors.

METHODS

Generating 2D virtual array sensor signals using a 3D convolutional neural network (CNN) requires huge computational costs and is impractical. Therefore, we installed virtual sensors between the real sensors along the spiral pattern in three different directions and used a 2D CNN to generate signals of the virtual sensors in each direction. Then we reconstructed a photoacoustic image using the signals from both the real sensors and the virtual sensors.

RESULTS

We evaluated the proposed method using simulation data and human palm measurement data. We found that these artifacts were significantly reduced in the images reconstructed using the proposed method, while the artifacts were strong in the images obtained only from the real sensor signals.

CONCLUSION

Using the proposed method, we were able to significantly reduce artifacts, and as a result, it became possible to recognize deep blood vessels. In addition, the processing time of the proposed method was sufficiently applicable to clinical measurement.

Malignancy diagnosis of liver lesion in contrast enhanced ultrasound using an end-to-end method based on deep learning

BACKGROUND

Contrast-enhanced ultrasound (CEUS) is considered as an efficient tool for focal liver lesion characterization, given it allows real-time scanning and provides dynamic tissue perfusion information. An accurate diagnosis of liver lesions with CEUS requires a precise interpretation of CEUS images. However,it is a highly experience dependent task which requires amount of training and practice. To help improve the constrains, this study aims to develop an end-to-end method based on deep learning to make malignancy diagnosis of liver lesions using CEUS.

METHODS

A total of 420 focal liver lesions with 136 benign cases and 284 malignant cases were included. A deep learning model based on a two-dimensional convolution neural network, a long short-term memory (LSTM), and a linear classifier (with sigmoid) was developed to analyze the CEUS loops from different contrast imaging phases. For comparison, a 3D-CNN based method and a machine-learning (ML)-based time-intensity curve (TIC) method were also implemented for performance evaluation.

RESULTS

Results of the 4-fold validation demonstrate that the mean AUC is 0.91, 0.88, and 0.78 for the proposed method, the 3D-CNN based method, and the ML-based TIC method, respectively.

CONCLUSIONS

The proposed CNN-LSTM method is promising in making malignancy diagnosis of liver lesions in CEUS without any additional manual features selection.

Applied deep learning in neurosurgery: identifying cerebrospinal fluid (CSF) shunt systems in hydrocephalus patients

BACKGROUND

Over the recent decades, the number of different manufacturers and models of cerebrospinal fluid shunt valves constantly increased. Proper identification of shunt valves on X-ray images is crucial to neurosurgeons and radiologists to derive further details of a specific shunt valve, such as opening pressure settings and MR scanning conditions. The main aim of this study is to evaluate the feasibility of an AI-assisted shunt valve detection system.

METHODS

The dataset used contains 2070 anonymized images of ten different, commonly used shunt valve types. All images were acquired from skull X-rays or scout CT-images. The images were randomly split into a 80% training and 20% validation set. An implementation in Python with the FastAi library was used to train a convolutional neural network (CNN) using a transfer learning method on a pre-trained model.

RESULTS

Overall, our model achieved an F1-score of 99% to predict the correct shunt valve model. F1-scores for individual shunt valves ranged from 92% for the Sophysa Sophy Mini SM8 to 100% for several other models.

CONCLUSION

This technology has the potential to automatically detect different shunt valve models in a fast and precise way and may facilitate the identification of an unknown shunt valve on X-ray or CT scout images. The deep learning model we developed could be integrated into PACS systems or standalone mobile applications to enhance clinical workflows.

Deep learning approach for discrimination of liver lesions using nine time-phase images of contrast-enhanced ultrasound

PURPOSE

Contrast-enhanced ultrasound (CEUS) shows different enhancement patterns depending on the time after administration of the contrast agent. The aim of this study was to evaluate the diagnostic performance of liver nodule characterization using our proposed deep learning model with input of nine CEUS images.

METHODS

A total of 181 liver lesions (48 benign, 78 hepatocellular carcinoma (HCC), and 55 non-HCC malignant) were included in this prospective study. CEUS were performed using the contrast agent Sonazoid, and in addition to B-mode images before injection, image clips were stored every minute up to 10 min. A deep learning model was developed by arranging three ResNet50 transfer learning models in parallel. This proposed model allowed inputting up to nine datasets of different phases of CEUS and performing image augmentation of nine images synchronously. Using the results, the correct prediction rate, sensitivity, and specificity between "benign" and "malignant" cases were analyzed for each combination of the time phase. These accuracy values were also compared with the washout score judged by a human.

RESULTS

The proposed model showed performance superior to the referential standard model when the dataset from B-mode to the 10-min images were used (sensitivity: 93.2%, specificity: 65.3%, average correct answer rate: 60.1%). It also maintained 90.2% sensitivity and 61.2% specificity even when the dataset was limited to 2 min after injection, and this accuracy was equivalent to or better than human scoring by experts.

CONCLUSION

Our proposed model has the potential to identify tumor types earlier than the Kupffer phase, but at the same time, machine learning confirmed that Kupffer-phase Sonazoid images contain essential information for the classification of liver nodules.

Diagnosis of focal liver lesions from ultrasound images using a pretrained residual neural network

OBJECTIVE

This study aims to develop a ResNet50-based deep learning model for focal liver lesion (FLL) classification in ultrasound images, comparing its performance with other models and prior research.

METHODOLOGY

We retrospectively collected 581 ultrasound images from the Chulabhorn Hospital's HCC surveillance and screening project (2010-2018). The dataset comprised five classes: non-FLL, hepatic cyst (Cyst), hemangioma (HMG), focal fatty sparing (FFS), and hepatocellular carcinoma (HCC). We conducted 5-fold cross-validation after random dataset partitioning, enhancing training data with data augmentation. Our models used modified pre-trained ResNet50, GGN, ResNet18, and VGG16 architectures. Model performance, assessed via confusion matrices for sensitivity, specificity, and accuracy, was compared across models and with prior studies.

RESULTS

ResNet50 outperformed other models, achieving a 5-fold cross-validation accuracy of 87 ± 2.2%. While VGG16 showed similar performance, it exhibited higher uncertainty. In the testing phase, the pretrained ResNet50 excelled in classifying non-FLL, cysts, and FFS. To compare with other research, ResNet50 surpassed the prior methods like two-layered feed-forward neural networks (FFNN) and CNN+ReLU in FLL diagnosis.

CONCLUSION

ResNet50 exhibited good performance in FLL diagnosis, especially for HCC classification, suggesting its potential for developing computer-aided FLL diagnosis. However, further refinement is required for HCC and HMG classification in future studies.

Machine learning for malignant versus benign focal liver lesions on US and CEUS: a meta-analysis

OBJECTIVES

To perform a meta-analysis of the diagnostic performance of learning (ML) algorithms (conventional and deep learning algorithms) for the classification of malignant versus benign focal liver lesions (FLLs) on US and CEUS.

METHODS

Available databases were searched for relevant published studies through September 2022. Studies met eligibility criteria if they evaluate the diagnostic performance of ML for the classification of malignant and benign focal liver lesions on US and CEUS. The pooled per-lesion sensitivities and specificities for each modality with 95% confidence intervals were calculated.

RESULTS

A total of 8 studies on US, 11 on CEUS, and 1 study evaluating both methods met the inclusion criteria with a total of 34,245 FLLs evaluated. The pooled sensitivity and specificity of ML for the malignancy classification of FLLs were 81.7% (95% CI, 77.2-85.4%) and 84.8% (95% CI, 76.0-90.8%) for US, compared to 87.1% (95% CI, 81.8-91.0%) and 87.0% (95% CI, 83.1-90.1%) for CEUS. In the subgroup analysis of studies that evaluated deep learning algorithms, the sensitivity and specificity of CEUS (n = 4) increased to 92.4% (95% CI, 88.5-95.0%) and 88.2% (95% CI, 81.1-92.9%).

CONCLUSIONS

The diagnostic performance of ML algorithms for the malignant classification of FLLs was high for both US and CEUS with overall similar sensitivity and specificity. The similar performance of US may be related to the higher prevalence of DL models in that group.

Artificial intelligence assists operators in real-time detection of focal liver lesions during ultrasound: A randomized controlled study

PURPOSE

Detection of hepatocellular carcinoma (HCC) is crucial during surveillance by ultrasound. We previously developed an artificial intelligence (AI) system based on convolutional neural network for detection of focal liver lesions (FLLs) in ultrasound. The primary aim of this study was to evaluate whether the AI system can assist non-expert operators to detect FLLs in real-time, during ultrasound examinations.

METHOD

This single-center prospective randomized controlled study evaluated the AI system in assisting non-expert and expert operators. Patients with and without FLLs were enrolled and had ultrasound performed twice, with and without AI assistance. McNemar's test was used to compare paired FLL detection rates and false positives between groups with and without AI assistance.

RESULTS

260 patients with 271 FLLs and 244 patients with 240 FLLs were enrolled into the groups of non-expert and expert operators, respectively. In non-experts, FLL detection rate in the AI assistance group was significantly higher than the no AI assistance group (36.9 % vs 21.4 %, p < 0.001). In experts, FLL detection rates were not significantly different between the groups with and without AI assistance (66.7 % vs 63.3 %, p = 0.32). False positive detection rates in the groups with and without AI assistance were not significantly different in both non-experts (14.2 % vs 9.2 %, p = 0.08) and experts (8.6 % vs 9.0 %, p = 0.85).

CONCLUSIONS

The AI system resulted in significant increase in detection of FLLs during ultrasound examinations by non-experts. Our findings may support future use of the AI system in resource-limited settings where ultrasound examinations are performed by non-experts. The study protocol was registered under the Thai Clinical Trial Registry (TCTR20201230003), which is part of the WHO ICTRP Registry Network. The registry can be accessed via the following URL: https://trialsearch.who.int/Trial2.aspx?TrialID=TCTR20201230003.

A study on the optimal condition of ground truth area for liver tumor detection in ultrasound images using deep learning

PURPOSE

In recent years, efforts to apply artificial intelligence (AI) to the medical field have been growing. In general, a vast amount of high-quality training data is necessary to make great AI. For tumor detection AI, annotation quality is important. In diagnosis and detection of tumors using ultrasound images, humans use not only the tumor area but also the surrounding information, such as the back echo of the tumor. Therefore, we investigated changes in detection accuracy when changing the size of the region of interest (ROI, ground truth area) relative to liver tumors in the training data for the detection AI.

METHODS

We defined D/L as the ratio of the maximum diameter (D) of the liver tumor to the ROI size (L). We created training data by changing the D/L value, and performed learning and testing with YOLOv3.

RESULTS

Our results showed that the detection accuracy was highest when the training data were created with a D/L ratio between 0.8 and 1.0. In other words, it was found that the detection accuracy was improved by setting the ground true bounding box for detection AI training to be in contact with the tumor or slightly larger. We also found that when the D/L ratio was distributed in the training data, the wider the distribution, the lower the detection accuracy.

CONCLUSIONS

Therefore, we recommend that the detector be trained with the D/L value close to a certain value between 0.8 and 1.0 for liver tumor detection from ultrasound images.

Improving artificial intelligence pipeline for liver malignancy diagnosis using ultrasound images and video frames

Recent developments of deep learning methods have demonstrated their feasibility in liver malignancy diagnosis using ultrasound (US) images. However, most of these methods require manual selection and annotation of US images by radiologists, which limit their practical application. On the other hand, US videos provide more comprehensive morphological information about liver masses and their relationships with surrounding structures than US images, potentially leading to a more accurate diagnosis. Here, we developed a fully automated artificial intelligence (AI) pipeline to imitate the workflow of radiologists for detecting liver masses and diagnosing liver malignancy. In this pipeline, we designed an automated mass-guided strategy that used segmentation information to direct diagnostic models to focus on liver masses, thus increasing diagnostic accuracy. The diagnostic models based on US videos utilized bi-directional convolutional long short-term memory modules with an attention-boosted module to learn and fuse spatiotemporal information from consecutive video frames. Using a large-scale dataset of 50 063 US images and video frames from 11 468 patients, we developed and tested the AI pipeline and investigated its applications. A dataset of annotated US images is available at https://doi.org/10.5281/zenodo.7272660.

Ensemble Learning of Multiple Models Using Deep Learning for Multiclass Classification of Ultrasound Images of Hepatic Masses

Ultrasound (US) is often used to diagnose liver masses. Ensemble learning has recently been commonly used for image classification, but its detailed methods are not fully optimized. The purpose of this study is to investigate the usefulness and comparison of some ensemble learning and ensemble pruning techniques using multiple convolutional neural network (CNN) trained models for image classification of liver masses in US images. Dataset of the US images were classified into four categories: benign liver tumor (BLT) 6320 images, liver cyst (LCY) 2320 images, metastatic liver cancer (MLC) 9720 images, primary liver cancer (PLC) 7840 images. In this study, 250 test images were randomly selected for each class, for a total of 1000 images, and the remaining images were used as the training. 16 different CNNs were used for training and testing ultrasound images. The ensemble learning used soft voting (SV), weighted average voting (WAV), weighted hard voting (WHV) and stacking (ST). All four types of ensemble learning (SV, ST, WAV, and WHV) showed higher values of accuracy than the single CNN. All four types also showed significantly higher deep learning (DL) performance than ResNeXt101 alone. For image classification of liver masses using US images, ensemble learning improved the performance of DL over a single CNN.

Artificial intelligence models for the diagnosis and management of liver diseases

With the development of more advanced methods for the diagnosis and treatment of diseases, the data required for medical care are becoming complex, and misinterpretation of information due to human error may result in serious consequences. Human error can be avoided with the support of artificial intelligence (AI). AI models trained with various medical data for diagnosis and management of liver diseases have been applied to hepatitis, fatty liver disease, liver cirrhosis, and liver cancer. Some of these models have been reported to outperform human experts in terms of performance, indicating their potential for supporting clinical practice given their high-speed output. This paper summarizes the recent advances in AI for liver disease and introduces the AI-aided diagnosis of liver tumors using B-mode ultrasonography.

Deep Learning for Approaching Hepatocellular Carcinoma Ultrasound Screening Dilemma: Identification of α-Fetoprotein-Negative Hepatocellular Carcinoma From Focal Liver Lesion Found in High-Risk Patients

Background

First-line surveillance on hepatitis B virus (HBV)-infected populations with B-mode ultrasound is relatively limited to identifying hepatocellular carcinoma (HCC) without elevated α-fetoprotein (AFP). To improve the present HCC surveillance strategy, the state of the art of artificial intelligence (AI), a deep learning (DL) approach, is proposed to assist in the diagnosis of a focal liver lesion (FLL) in HBV-infected liver background.

Methods

Our proposed deep learning model was based on B-mode ultrasound images of surgery that proved 209 HCC and 198 focal nodular hyperplasia (FNH) cases with 413 lesions. The model cohort and test cohort were set at a ratio of 3:1, in which the test cohort was composed of AFP-negative HBV-infected cases. Four additional deep learning models (MobileNet, Resnet50, DenseNet121, and InceptionV3) were also constructed as comparative baselines. To evaluate the models in terms of diagnostic power, sensitivity, specificity, accuracy, confusion matrix, F1-score, and area under the receiver operating characteristic curve (AUC) were calculated in the test cohort.

Results

The AUC of our model, Xception, achieved 93.68% in the test cohort, superior to other baselines (89.06%, 85.67%, 83.94%, and 78.13% respectively for MobileNet, Resnet50, DenseNet121, and InceptionV3). In terms of diagnostic power, our model showed sensitivity, specificity, accuracy, and F1-score of 96.08%, 76.92%, 86.41%, and 87.50%, respectively, and PPV, NPV, FPR, and FNR calculated from the confusion matrix were respectively 80.33%, 95.24%, 23.08%, and 3.92% in identifying AFP-negative HCC from HBV-infected FLL cases. Satisfactory robustness of our proposed model was shown based on 5-fold cross-validation performed among the models above.

Conclusions

Our DL approach has great potential to assist B-mode ultrasound in identifying AFP-negative HCC from FLL found in surveillance of HBV-infected patients.