A multicenter clinical AI system study for detection and diagnosis of focal liver lesions

Early and accurate diagnosis of focal liver lesions is crucial for effective treatment and prognosis. We developed and validated a fully automated diagnostic system named Liver Artificial Intelligence Diagnosis System (LiAIDS) based on a diverse sample of 12,610 patients from 18 hospitals, both retrospectively and prospectively. In this study, LiAIDS achieved an F1-score of 0.940 for benign and 0.692 for malignant lesions, outperforming junior radiologists (benign: 0.830-0.890, malignant: 0.230-0.360) and being on par with senior radiologists (benign: 0.920-0.950, malignant: 0.550-0.650). Furthermore, with the assistance of LiAIDS, the diagnostic accuracy of all radiologists improved. For benign and malignant lesions, junior radiologists’ F1-scores improved to 0.936-0.946 and 0.667-0.680 respectively, while seniors improved to 0.950-0.961 and 0.679-0.753. Additionally, in a triage study of 13,192 consecutive patients, LiAIDS automatically classified 76.46% of patients as low risk with a high NPV of 99.0%. The evidence suggests that LiAIDS can serve as a routine diagnostic tool and enhance the diagnostic capabilities of radiologists for liver lesions.

experimental group/condition, given as a discrete number and unit of measurement rements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided solely by name; describe more complex techniques in the Methods section.
or corrections, such as tests of normality and adjustment for multiple comparisons including central tendency (e.g.means) or other basic estimates AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g.confidence intervals) (e.g.F,t,r) with confidence intervals, effect sizes, degrees of freedom choice of priors and Markov chain Monte Carlo settings of the appropriate level for tests and full reporting of outcomes , Pearson's r), indicating how they were calculated Our web collection onstatistics for biologistscontains articles on many of the points above.code For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors andreviewers.We strongly encourage code deposition in a community repository (e.g.GitHub).See the Nature Portfolio guidelines for submitting code & software statement.This statement should provide the following information, publicly available datasets that the statement adheres to our policy.
No software was used for data collection.
The architecture of our system is an integration of innovative technologies, including an enhanced Faster R-CNN with the CSwin Transformer for Net augmented with the scSE attention mechanism, and DenseNet.Policy information about studies with human participants or human data.See also policy information about sex, gender (identity/presentation), and sexual orientation and race, ethnicity and racism.

Reporting on sex and gender
Reporting on race, ethnicity, or other socially relevant groupings

Ethics oversight
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Field-specific reporting
Please select the one below that is the best fit for your research.If you are not sure, read the appropriate sections before making your selection.

Lifesciences Behavioural & social sciences Ecological, evolutionary & environmental science
The design of this study took into account both sex and/or gender, with these attributes determined based onself-reporting at the time of patient enrollment.
This study does not involve any socially constructed or socially relevant categorization variable(s), including but not limited to race, ethnicity, or other social relevant groupings.
This study has received approval from the Institutional Review Board (IRB) of Sir Run Run Shaw Hospital (SRRSH) and was carried out in adherence to the Declaration of Helsinki.Additionally, the prospective component of this study is officially registered with the Chinese Clinical Trial Registry, under the identifier ChiCTR2100045278 (accessible at [https://www.chictr.org.cn/showproj.html?proj=124700], registration date: April 10, 2021).In parallel, all 17 collaborating institutions obtained requisite IRB approvals for their participation in the retrospective aspects of the study.Owing to its noninvasive methodology, the IRB granted a waiver for the informed consent requirement.
In our study, we enrolled participants aged 14 years and above who underwent enhanced CT scans for diagnosing focal liver lesions (FLLs).Each participant underwent a comprehensive triple-phase CT scan, which included non-contrast, arterial, and portal venous phases.We also collected key clinical data, encompassing basic demographics like age and gender, along with relevant medical history such as hepatitis, cirrhosis, cholangiolithiasis, and extra-hepatic tumors.This information was sourced from self-reports and Electronic Medical Records (EMRs) at the time of enrollment.To mitigate potential biases and enhance the representativeness of our training dataset, we diversified our sample collection by including participants from 18 different hospitals.This approach ensured a broad spectrum of demographic and clinical variations, contributing to the robustness and generalizability of our findings.
This study recruited participants aged 14 years and older who underwent enhanced CT scans for focal liver lesions (FLLs).
Patients who had undergone any form of treatment for FLLs prior to the contrast enhanced CT scan were excluded.This includes those who received surgery, transcatheter arterial chemoembolization (TACE), radiofrequency ablation, chemotherapy, radiotherapy, targeted drug therapy or immunotherapy.

April2023
For a reference copy of the document with all sections, see nature.

Plants
This study included retrospective data to develop and validate the proposed LiAIDS.In addition, a prospective study further include July1st, 2020 and June 30th, 2021.Furthermore, 13,192 consecutive patients admitted to SRRSH between May 1st 2022 and August 2022were collected for a study of patient triage.As a common rule of thumb in trad have at least 30 instances for each class in the test set.Although somewhat arbitrary, this guideline typically strikes a de achieving statistical significance and managing compu and the counts for each of the 7 lesion classes all exceed this threshold, as detailed in Table 1.Therefore, we are confiden lesions in each class is ample for reliable model assessment.As for our non exceeded our initial sample size calculat (alpha, α) of 0.05, a power (1-beta, 1-β) of 0.8, and an expected efficacy accuracy of 0.92, had anticipated a need for 173 participants.Therefore, our study was adequately powered as the actual number of recruited participants surpassed the calculated necessity.
Exclusion criteria were: (1) patients who received any form of treatment for FLLs prior to the contrast transcatheter arterial chemoembolization (TACE), radiofrequency ablation, chemotherapy, radiotherapy, targeted drug therapy, immunotherapy; (2) patients who had a clinical diagnosis of malignant lesions but lacked pathological confirmation; (3) be lacked both a histopathological report and a consensus agreement; (4) cases with compromised CT image quality due to reasons patient movement, incorrect positioning, information, including basic patient data (for example age and gender) and relevant medical history (such as hepatitis, cirrh cholangiolithiasis, and extra-hepatic tumors).
In this study, the retrospective dataset was partitioned into an internal cohort comprising data from 15 hospitals and thre external cohorts, consisting of data from the remaining 3 hospitals, respectively.Participants in the internal cohort were r internal training and validation cohorts at a 4:1 ratio.For the purpose of external validation, excluding the largest SRRSH, to form our external validation cohorts.This approach guarantees a comprehensive performance as a wide range of data, thereby capturing a greater diversity of pat materials, systems and methods materials, experimental systems and methods used in many studies.Here, ystem or method listed is relevant to your study.If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.Method data from 11,385 patients managed in18hospitals in China between January to develop and validate the proposed LiAIDS.In addition, a prospective study further included 1,225 patients treated at SRRSH between July1st, 2020 andJune 30th, 2021. Furthermore, 13,192 consecutive patients admitted to SRRSH between May 1st 2022 and August 2022were collected for a study of patient triage.As a common rule of thumb in traditional statistical studies, it's usually recommended to have at least 30 instances for each class in the test set.Although somewhat arbitrary, this guideline typically strikes a de achieving statistical significance and managing computational expense.In our study, we use a large number of samples from different centers and the counts for each of the 7 lesion classes all exceed this threshold, as detailed in Table 1.Therefore, we are confiden is ample for reliable model assessment.As for our non-inferiority study, the 183 pathologically confirmed patients exceeded our initial sample size calculation for the trial.This estimation, based on a non-inferiority margin (delta, δ) of 0.1, a significance level β) of 0.8, and an expected efficacy accuracy of 0.92, had anticipated a need for 173 participants.e, our study was adequately powered as the actual number of recruited participants surpassed the calculated necessity.
Exclusion criteria were: (1) patients who received any form of treatment for FLLs prior to the contrast-enhanced CT scan, including surg transcatheter arterial chemoembolization (TACE), radiofrequency ablation, chemotherapy, radiotherapy, targeted drug therapy, immunotherapy; (2) patients who had a clinical diagnosis of malignant lesions but lacked pathological confirmation; (3) be lacked both a histopathological report and a consensus agreement; (4) cases with compromised CT image quality due to reasons positioning, presence of metallic objects, or equipment malfunctions; and (5) information, including basic patient data (for example age and gender) and relevant medical history (such as hepatitis, cirrh tumors).
he retrospective dataset was partitioned into an internal cohort comprising data from 15 hospitals and thre external cohorts, consisting of data from the remaining 3 hospitals, respectively.Participants in the internal cohort were r internal training and validation cohorts at a 4:1 ratio.For the purpose of external validation, we specifically selected the three largest sites, excluding the largest SRRSH, to form our external validation cohorts.This approach guarantees a comprehensive performance as a wide range of data, thereby capturing a greater diversity of patient scenarios.group allocation during data collection and analysis.
The data are available from the corresponding author upon reasonable request.The architecture of our system is an integration of innovative  January 1st, 2010and June 30th 2020, d 1,225 patients treated at SRRSH between July1st, 2020and June 30th, 2021. Furthermore, 13,192 consecutive patients admitted to SRRSH between May 1st 2022 and August 31th itional statistical studies, it's usually recommended to have at least 30 instances for each class in the test set.Although somewhat arbitrary, this guideline typically strikes a decent balance between tational expense.In our study, we use a large number of samples from different centers and the counts for each of the 7 lesion classes all exceed this threshold, as detailed in Table 1.Therefore, we are confident that the number of inferiority study, the 183 pathologically confirmed patients inferiority margin (delta, δ) of 0.1, a significance level β) of 0.8, and an expected efficacy accuracy of 0.92, had anticipated a need for 173 participants.e, our study was adequately powered as the actual number of recruited participants surpassed the calculated necessity.enhanced CT scan, including surgery, transcatheter arterial chemoembolization (TACE), radiofrequency ablation, chemotherapy, radiotherapy, targeted drug therapy, and immunotherapy; (2) patients who had a clinical diagnosis of malignant lesions but lacked pathological confirmation; (3) benign cases that lacked both a histopathological report and a consensus agreement; (4) cases with compromised CT image quality due to reasons including (5) cases that lacked essential clinical information, including basic patient data (for example age and gender) and relevant medical history (such as hepatitis, cirrhosis, he retrospective dataset was partitioned into an internal cohort comprising data from 15 hospitals and three independent external cohorts, consisting of data from the remaining 3 hospitals, respectively.Participants in the internal cohort were randomly divided into we specifically selected the three largest sites, excluding the largest SRRSH, to form our external validation cohorts.This approach guarantees a comprehensive performance assessment over 3 natureportfolio|reportingsummary April2023 onding author upon reasonable request.The architecture of our system is an integration of innovative Net augmented with the scSE nd DenseNet.For transparency and reproducibility, the source code and models for each component are available CNN is hosted at https://github.com/open-Transformer;U-Net is available at UNet; the scSE Attention Unit is located at https://gitcode.net/mirrors/shanglianlm0525/pytorchr/Attention/SEvariants.py; and DenseNet can be accessed at https://github.com/xmuyzz/3D-CNN- com/documents/nr-reporting-summary-flat.pdf n/a Involved in the Life sciences study design All studies must disclose on these points even when the materials, We require information from authors about some types of materials, system or method listed is relevant to your study.If you are not sure if a list item applies to your research, read the appro Materials For transparency and r and models for each component are available through open-source platforms.The respective repositories are as follows: Faster R mmlab/mmdetection; CSwin Transformer can be found at https://github.com/microsoft/CSWinavailable at https://github.com/milesial/Pytorch-UNet; the scSE Attention Unit is located at https://gitcode.net/mirrors/shanglianlm0525/pytorch/blob/master/Attention/SEvariants.py; and DenseNet can be accessed at https://github.com/xmuyzz/3DPyTorch/blob/master/models/DenseNet.py.All analytical data underpinning the findings of this study are incorporated within this paper in the designated Source Data files (Source_data_Figure_3.xlsx to Source_data_Figure_7.xlsx and Source_data_Figure_S1.xlsx to Source_data_Figure_S3.xlsx).The raw imaging and clinical dataset being anonymized.These datasets can be obtained from the corresponding author upon reasonable request.
technologies, including an enhanced Faster R-CNN with the CSwin Transformer for feature extraction, a U-Net augmented with the scSE nd DenseNet.For transparency and reproducibility, the source code and models for each component are available source platforms.The respective repositories are as follows: Faster R-CNN is hosted at https://github.com/openSwin Transformer can be found at https://github.com/microsoft/CSWin-Transformer;U UNet; the scSE Attention Unit is located at https://gitcode.net/mirrors/shanglianlm0525/pytorchr/Attention/SEvariants.py; and DenseNet can be accessed at https://github.com/xmuyzz/3DPyTorch/blob/master/models/DenseNet.py.
methodsHere, indicate whether each material, priate section before selecting a response.