External Test Set Research Articles

Identification of Biomarkers for Endometrial Cancer Based on Natural Killer Cell-Related Genes

Endometrial cancer (EC) is a malignant tumor. Natural killer (NK) cells play a crucial role in various cancers, but their role in EC is unclear. To this purpose, in this paper, differential expression analysis was performed on transcriptome data from the TCGA database, and the obtained DEGs and the collected NRGs were intersected, and single-factor Cox regression analysis and Lasso-Cox regression analysis were performed on the intersected genes to obtain prognosis-related genes and risk model, respectively. These genes and models were validated by Kaplan-Meier (KM) survival curve analysis and ROC analysis on the internal and external test sets. In addition, nomogram models were constructed based on prognosis-associated genes, sample risk scores, and clinical factors. Finally, we explored the immune landscape of high- and low-risk groups of EC. The results showed that the risk models constructed in this paper exhibited excellent predictive effects, which will facilitate research on the precision treatment of EC. There were significant differences in prognosis, immune cell infiltration abundance, immune checkpoint-associated genes, and HLA gene expression between high- and low-risk groups of EC. The risk model in this paper can provide a reference for the personalized treatment of EC. In addition, we performed RT-qPCR to validate the levels of genes significantly associated with prognosis.

Use of Pretreatment Perfusion MRI-based Intratumoral Heterogeneity to Predict Pathologic Response of Triple-Negative Breast Cancer to Neoadjuvant Chemoimmunotherapy.

Background Neoadjuvant chemoimmunotherapy (NACI) has significantly increased the rate of pathologic complete response (pCR) in patients with early-stage triple-negative breast cancer (TNBC), although predictors of response to this regimen have not been identified. Purpose To investigate pretreatment perfusion MRI-based radiomics as a predictive marker for pCR in patients with TNBC undergoing NACI. Materials and Methods This prospective study enrolled women with early-stage TNBC who underwent NACI at two different centers from August 2021 to July 2023. Pretreatment dynamic contrast-enhanced MRI scans obtained using scanners from multiple vendors were analyzed using the Tofts model to segment tumors and analyze pharmacokinetic parameters. Radiomics features were extracted from the rate constant for contrast agent plasma-to-interstitial transfer (or Ktrans), volume fraction of extravascular and extracellular space (Ve), and maximum contrast agent uptake rate (Slopemax) maps and analyzed using unsupervised correlation and least absolute shrinkage and selector operator, or LASSO, to develop a radiomics score. Score effectiveness was assessed using the area under the receiver operating characteristic curve (AUC), and multivariable logistic regression was used to develop a multimodal nomogram for enhanced prediction. The discrimination, calibration, and clinical utility of the nomogram were evaluated in an external test set. Results The training set included 112 female participants from center 1 (mean age, 52 years ± 11 [SD]), and the external test set included 83 female participants from center 2 (mean age, 47 years ± 11). The radiomics score demonstrated an AUC of 0.80 (95% CI: 0.70, 0.89) for predicting pCR. A nomogram incorporating the radiomics score, grade, and Ki-67 yielded an AUC of 0.86 (95% CI: 0.78, 0.94) in the test set. Associations were found between higher radiomics score (>0.25) and tumor size (P < .001), washout enhancement (P = .01), androgen receptor expression (P = .009), and programmed death ligand 1 expression (P = .01), demonstrating a correlation with tumor immune environment in participants with TNBC. Conclusion A radiomics score derived from pharmacokinetic parameters at pretreatment dynamic contrast-enhanced MRI exhibited good performance for predicting pCR in participants with TNBC undergoing NACI, and could potentially be used to enhance clinical decision making. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Rauch in this issue.

Deep learning-based structure segmentation and intramuscular fat annotation on lumbar magnetic resonance imaging.

Lumbar disc herniation (LDH) is a prevalent cause of low back pain. LDH patients commonly experience paraspinal muscle atrophy and fatty infiltration (FI), which further exacerbates the symptoms of low back pain. Magnetic resonance imaging (MRI) is crucial for assessing paraspinal muscle condition. Our study aims to develop a dual-model for automated muscle segmentation and FI annotation on MRI, assisting clinicians evaluate LDH conditions comprehensively. The study retrospectively collected data diagnosed with LDH from December 2020 to May 2022. The dataset was split into a 7:3 ratio for training and testing, with an external test set prepared to validate model generalizability. The model's performance was evaluated using average precision (AP), recall and F1 score. The consistency was assessed using the Dice similarity coefficient (DSC) and Cohen's Kappa. The mean absolute percentage error (MAPE) was calculated to assess the error of the model measurements of relative cross-sectional area (rCSA) and FI. Calculate the MAPE of FI measured by threshold algorithms to compare with the model. A total of 417 patients being evaluated, comprising 216 males and 201 females, with a mean age of 49 ± 15 years. In the internal test set, the muscle segmentation model achieved an overall DSC of 0.92 ± 0.10, recall of 92.60%, and AP of 0.98. The fat annotation model attained a recall of 91.30%, F1 Score of 0.82, and Cohen's Kappa of 0.76. However, there was a decrease on the external test set. For rCSA measurements, except for longissimus (10.89%), the MAPE of other muscles was less than 10%. When comparing the errors of FI for each paraspinal muscle, the MAPE of the model was lower than that of the threshold algorithm. The models demonstrate outstanding performance, with lower error in FI measurement compared to thresholding algorithms.

Improving performance in colorectal cancer histology decomposition using deep and ensemble machine learning

In routine colorectal cancer management, histologic samples stained with hematoxylin and eosin are commonly used. Nonetheless, their potential for defining objective biomarkers for patient stratification and treatment selection is still being explored. The current gold standard relies on expensive and time-consuming genetic tests. However, recent research highlights the potential of convolutional neural networks (CNNs) to facilitate the extraction of clinically relevant biomarkers from these readily available images. These CNN-based biomarkers can predict patient outcomes comparably to golden standards, with the added advantages of speed, automation, and minimal cost. The predictive potential of CNN-based biomarkers fundamentally relies on the ability of CNNs to accurately classify diverse tissue types from whole slide microscope images. Consequently, enhancing the accuracy of tissue class decomposition is critical to amplifying the prognostic potential of imaging-based biomarkers. This study introduces a hybrid deep transfer learning and ensemble machine learning model that improves upon previous approaches, including a transformer and neural architecture search baseline for this task. We employed a pairing of the EfficientNetV2 architecture with a random forest classification head. Our model achieved 96.74% accuracy (95% CI: 96.3%-97.1%) on the external test set and 99.89% on the internal test set. Recognizing the potential of these models in the task, we have made them publicly available.

RawECGNet: Deep Learning Generalization for Atrial Fibrillation Detection From the Raw ECG.

Deep learning models for detecting episodes of atrial fibrillation (AF) using rhythm information in long-term ambulatory ECG recordings have shown high performance. However, the rhythm-based approach does not take advantage of the morphological information conveyed by the different ECG waveforms, particularly the f-waves. As a result, the performance of such models may be inherently limited. To address this limitation, we have developed a deep learning model, named RawECGNet, to detect episodes of AF and atrial flutter (AFl) using the raw, single-lead ECG. We compare the generalization performance of RawECGNet on two external data sets that account for distribution shifts in geography, ethnicity, and lead position. RawECGNet is further benchmarked against a state-of-the-art deep learning model, named ArNet2, which utilizes rhythm information as input. Using RawECGNet, the results for the different leads in the external test sets in terms of the F1 score were 0.91-0.94 in RBDB and 0.93 in SHDB, compared to 0.89-0.91 in RBDB and 0.91 in SHDB for ArNet2. The results highlight RawECGNet as a high-performance, generalizable algorithm for detection of AF and AFl episodes, exploiting information on both rhythm and morphology.

Deep Learning Segmentation of Infiltrative and Enhancing Cellular Tumor at Pre- and Posttreatment Multishell Diffusion MRI of Glioblastoma.

Purpose To develop and validate a deep learning (DL) method to detect and segment enhancing and nonenhancing cellular tumor on pre- and posttreatment MRI scans in patients with glioblastoma and to predict overall survival (OS) and progression-free survival (PFS). Materials and Methods This retrospective study included 1397 MRI scans in 1297 patients with glioblastoma, including an internal set of 243 MRI scans (January 2010 to June 2022) for model training and cross-validation and four external test cohorts. Cellular tumor maps were segmented by two radiologists on the basis of imaging, clinical history, and pathologic findings. Multimodal MRI data with perfusion and multishell diffusion imaging were inputted into a nnU-Net DL model to segment cellular tumor. Segmentation performance (Dice score) and performance in distinguishing recurrent tumor from posttreatment changes (area under the receiver operating characteristic curve [AUC]) were quantified. Model performance in predicting OS and PFS was assessed using Cox multivariable analysis. Results A cohort of 178 patients (mean age, 56 years ± 13 [SD]; 116 male, 62 female) with 243 MRI timepoints, as well as four external datasets with 55, 70, 610, and 419 MRI timepoints, respectively, were evaluated. The median Dice score was 0.79 (IQR, 0.53-0.89), and the AUC for detecting residual or recurrent tumor was 0.84 (95% CI: 0.79, 0.89). In the internal test set, estimated cellular tumor volume was significantly associated with OS (hazard ratio [HR] = 1.04 per milliliter; P < .001) and PFS (HR = 1.04 per milliliter; P < .001) after adjustment for age, sex, and gross total resection (GTR) status. In the external test sets, estimated cellular tumor volume was significantly associated with OS (HR = 1.01 per milliliter; P < .001) after adjustment for age, sex, and GTR status. Conclusion A DL model incorporating advanced imaging could accurately segment enhancing and nonenhancing cellular tumor, distinguish recurrent or residual tumor from posttreatment changes, and predict OS and PFS in patients with glioblastoma. Keywords: Segmentation, Glioblastoma, Multishell Diffusion MRI Supplemental material is available for this article. © RSNA, 2024.

Artificial Intelligence-Based Surgery Support Model Using Intraoperative Radiographs for Assessing the Acetabular Component Angle

BackgroundThis study aimed to develop an artificial intelligence–based surgical support model for assessing the acetabular component angle using intraoperative radiographs during total hip arthroplasty and verify its accuracy. MethodsA total of 268 hips were analyzed. At first, 268 preoperative and intraoperative anteroposterior pelvic radiographs were amplified to 536. These radiographs were used to create a learning model to estimate the acetabular component angle from the radiographs intraoperatively. The ground truth was the anteversion and inclination angles obtained from the computed tomography–based navigation system intraoperatively. Bone landmarks on the preoperative and intraoperative radiographs were manually annotated. The distances and angles between each landmark were used as predictor variables. The estimation accuracy was assessed for internal and external test datasets. Mean absolute error (MAE) and R2 values were used as accuracy measures. ResultsThe MAE and R2 for the internal test set showed 2.19 and 0.850 for anteversion, and 1.18 and 0.805 for inclination, respectively. The MAE and R2 for the external test set showed 2.78 and 0.789 for anteversion, and 1.56 and 0.744 for inclination, respectively. ConclusionsWe developed an artificial intelligence–based surgical support model for accurately assessing the acetabular component angle using intraoperative radiographs. Excellent estimation accuracy was confirmed for the external test set. In the future, the model may help to reduce the risk of adverse postoperative events.

Improving 18F-FDG PET Quantification Through a Spatial Normalization Method.

Quantification of 18F-FDG PET images is useful for accurate diagnosis and evaluation of various brain diseases, including brain tumors, epilepsy, dementia, and Parkinson disease. However, accurate quantification of 18F-FDG PET images requires matched 3-dimensional T1 MRI scans of the same individuals to provide detailed information on brain anatomy. In this paper, we propose a transfer learning approach to adapt a pretrained deep neural network model from amyloid PET to spatially normalize 18F-FDG PET images without the need for 3-dimensional MRI. Methods: The proposed method is based on a deep learning model for automatic spatial normalization of 18F-FDG brain PET images, which was developed by fine-tuning a pretrained model for amyloid PET using only 103 18F-FDG PET and MR images. After training, the algorithm was tested on 65 internal and 78 external test sets. All T1 MR images with a 1-mm isotropic voxel size were processed with FreeSurfer software to provide cortical segmentation maps used to extract a ground-truth regional SUV ratio using cerebellar gray matter as a reference region. These values were compared with those from spatial normalization-based quantification methods using the proposed method and statistical parametric mapping software. Results: The proposed method showed superior spatial normalization compared with statistical parametric mapping, as evidenced by increased normalized mutual information and better size and shape matching in PET images. Quantitative evaluation revealed a consistently higher SUV ratio correlation and intraclass correlation coefficients for the proposed method across various brain regions in both internal and external datasets. The remarkably good correlation and intraclass correlation coefficient values of the proposed method for the external dataset are noteworthy, considering the dataset's different ethnic distribution and the use of different PET scanners and image reconstruction algorithms. Conclusion: This study successfully applied transfer learning to a deep neural network for 18F-FDG PET spatial normalization, demonstrating its resource efficiency and improved performance. This highlights the efficacy of transfer learning, which requires a smaller number of datasets than does the original network training, thus increasing the potential for broader use of deep learning-based brain PET spatial normalization techniques for various clinical and research radiotracers.

Precise Image-level Localization of Intracranial Hemorrhage on Head CT Scans with Deep Learning Models Trained on Study-level Labels.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop a highly generalizable weakly supervised model to automatically detect and localize image- level intracranial hemorrhage (ICH) using study-level labels. Materials and Methods In this retrospective study, the proposed model was pretrained on the image-level RSNA dataset and fine-tuned on a local dataset using attention-based bidirectional long-short-term memory networks. This local training dataset included 10,699 noncontrast head CT scans from 7469 patients with ICH study-level labels extracted from radiology reports. Model performance was compared with that of two senior neuroradiologists on 100 random test scans using the McNemar test, and its generalizability was evaluated on an external independent dataset. Results The model achieved a positive predictive value (PPV) of 85.7% (95% CI: [84.0%, 87.4%]) and an AUC of 0.96 (95% CI: [0.96, 0.97]) on the held-out local test set (n = 7243, 3721 female) and 89.3% (95% CI: [87.8%, 90.7%]) and 0.96 (95% CI: [0.96, 0.97]), respectively, on the external test set (n = 491, 178 female). For 100 randomly selected samples, the model achieved performance on par with two neuroradiologists, but with a significantly faster (P < .05) diagnostic time of 5.04 seconds per scan (versus 86 seconds and 22.2 seconds for the two neuroradiologists, respectively). The model's attention weights and heatmaps visually aligned with neuroradiologists' interpretations. Conclusion The proposed model demonstrated high generalizability and high PPVs, offering a valuable tool for expedited ICH detection and prioritization while reducing false-positive interruptions in radiologists' workflows. ©RSNA, 2024.

Exploring and identifying the imaging biomarkers for predicting anti-VEGF treatment response in polypoidal choroidal vasculopathy: aprospective multicenter study.

Polypoidal choroidal vasculopathy (PCV) is a hemorrhagic fundus disease that can lead to permanent vision loss. Predicting the treatment response to anti-VEGF monotherapy in PCV is consistently challenging. We aimed to conduct a prospective multicenter study to explore and identify the imaging biomarkers for predicting the anti-VEGF treatment response in PCV patients, establish predictive model, and undergo multicenter validation. This prospective multicenter study utilized clinical characteristics and images of treatment naïve PCV patients from 15 ophthalmic centers nationwide to screen biomarkers, develop model, and validate its performance. Patients from Peking Union Medical College Hospital were randomly divided into a training set and an internal validation set. A nomogram was established by univariate, LASSO regression, and multivariate regression analysis. Patients from the other 14 centers served as an external test set. Area under the curve (AUC), sensitivity, specificity, and accuracy were calculated. Decision curve analysis (DCA) and clinical impact curve (CIC) were utilized to evaluate the practical utility in clinical decision-making. The eye distribution for the training set, internal validation set, and external test set were 66, 31, and 71, respectively. The 'Good responder' exhibited a thinner subfoveal choroidal thickness (SFCT) (230.67 ± 61.96 vs. 314.42 ± 88.00 μm, p < 0.001), lower choroidal vascularity index (CVI) (0.31 ± 0.08 vs. 0.36 ± 0.05, p = 0.006), fewer choroidal vascular hyperpermeability (CVH) (31.0 vs. 62.2%, p = 0.012), and more intraretinal fluid (IRF) (58.6 vs. 29.7%, p = 0.018). SFCT (OR 0.990; 95% CI 0.981-0.999; p = 0.033) and CVI (OR 0.844; 95% CI 0.732-0.971; p = 0.018) were ultimately included as the optimal predictive biomarkers and presented in the form of a nomogram. The model demonstrated AUC of 0.837 (95% CI 0.738-0.936), 0.891 (95% CI 0.765-1.000), and 0.901 (95% CI 0.824-0.978) for predicting 'Good responder' in the training set, internal validation set, and external test set, respectively, with excellent sensitivity, specificity, and practical utility. Thinner SFCT and lower CVI can serve as imaging biomarkers for predicting good treatment response to anti-VEGF monotherapy in PCV patients. The nomogram based on these biomarkers exhibited satisfactory performances.

Multiregional dynamic contrast-enhanced MRI-based integrated system for predicting pathological complete response of axillary lymph node to neoadjuvant chemotherapy in breast cancer: multicentre study

The accurate evaluation of axillary lymph node (ALN) response to neoadjuvant chemotherapy (NAC) in breast cancer holds great value. This study aimed to develop an artificial intelligence system utilising multiregional dynamic contrast-enhanced MRI (DCE-MRI) and clinicopathological characteristics to predict axillary pathological complete response (pCR) after NAC in breast cancer. This study included retrospective and prospective datasets from six medical centres in China between May 2018 and December 2023. A fully automated integrated system based on deep learning (FAIS-DL) was built to perform tumour and ALN segmentation and axillary pCR prediction sequentially. The predictive performance of FAIS-DL was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. RNA sequencing analysis were conducted on 45 patients to explore the biological basis of FAIS-DL. 1145 patients (mean age, 50 years±10 [SD]) were evaluated. Among these patients, 506 were in the training and validation sets (axillary pCR rate of 40.3%), 127 in the internal test set (axillary pCR rate of 37.8%), 414 in the pooled external test set (axillary pCR rate of 48.8%), and 98 in the prospective test set (axillary pCR rate of 43.9%). For predicting axillary pCR, FAIS-DL achieved AUCs of 0.95, 0.93, and 0.94 in the internal test set, pooled external test set, and prospective test set, respectively, which were also significantly higher than those of the clinical model and deep learning models based on single-regional DCE-MRI (all P<0.05, DeLong test). In the pooled external and prospective test sets, the FAIS-DL decreased the unnecessary axillary lymph node dissection rate from 47.9% to 6.8%, and increased the benefit rate from 52.2% to 86.5%. RNA sequencing analysis revealed that high FAIS-DL scores were associated with the upregulation of immune-mediated genes and pathways. FAIS-DL has demonstrated satisfactory performance in predicting axillary pCR, which may guide the formulation of personalised treatment regimens for patients with breast cancer in clinical practice. This study was supported by the National Natural Science Foundation of China (82371933), National Natural Science Foundation of Shandong Province of China (ZR2021MH120), Mount Taishan Scholars and Young Experts Program (tsqn202211378), Key Projects of China Medicine Education Association (2022KTM030), China Postdoctoral Science Foundation (314730), and Beijing Postdoctoral Research Foundation (2023-zz-012).

Tumor contour irregularity on preoperative CT predicts prognosis in renal cell carcinoma: a multi-institutional study

Radiology-based prognostic biomarkers play a crucial role in patient counseling, enhancing surveillance, and designing clinical trials effectively. This study aims to assess the predictive significance of preoperative CT-based tumor contour irregularity in determining clinical outcomes among patients with renal cell carcinoma (RCC). We conducted a retrospective multi-institutional review involving 2218 patients pathologically diagnosed with RCC. The training and internal validation sets included patients at Zhongshan Hospital between January 2009 and August 2019. The external test set comprised patients from the First Affiliated Hospital, Zhejiang University School of Medicine (January 2016 to January 2018), the Xiamen Branch of Zhongshan Hospital (November 2017 to June 2023), and the Cancer Imaging Archive. The contour irregularity degree (CID), quantified as the ratio of irregular cross-sections to the total tumor cross-sections, was analyzed for its prognostic relevance across different subgroups of RCC patients. A novel CID-based scoring system was developed, and its predictive efficacy was evaluated and compared with existing prognostic models. The CID exhibited significant discriminatory power in predicting overall survival (OS), recurrence-free survival (RFS), and disease-specific survival (DSS) among patients with RCC tumors measuring 3cm or larger (all p<0.001). Multivariate analyses confirmed the CID as an independent prognostic indicator. Notably, the CID augmented prognostic stratification among RCC patients within distinct risk subgroups delineated by SSIGN models and ISUP grades. The CID-based nomogram (C-Model) demonstrated robust predictive performance, with C-index values of 0.88 (95%CI: 0.84-0.92) in the training set, 0.92 (95%CI: 0.88-0.98) in the internal validation set, and 0.86 (95%CI: 0.81-0.90) in the external test set, surpassing existing prognostic models. Routine imaging-based assessment of the CID serves as an independent prognostic factor, offering incremental prognostic value to existing models in RCC patients with tumors measuring 3cm or larger. This study was funded by grants from National Natural Science Foundation of China; Shanghai Municipal Health Commission; China National Key R&D Program and Science and Technology Commission of Shanghai Municipality.

Deep Learning Radiomics Features of Mediastinal Fat and Pulmonary Nodules on Lung CT Images Distinguish Benignancy and Malignancy.

This study investigated the relationship between mediastinal fat and pulmonary nodule status, aiming to develop a deep learning-based radiomics model for diagnosing benign and malignant pulmonary nodules. We proposed a combined model using CT images of both pulmonary nodules and the fat around the chest (mediastinal fat). Patients from three centers were divided into training, validation, internal testing, and external testing sets. Quantitative radiomics and deep learning features from CT images served as predictive factors. A logistic regression model was used to combine data from both pulmonary nodules and mediastinal adipose regions, and personalized nomograms were created to evaluate the predictive performance. The model incorporating mediastinal fat outperformed the nodule-only model, with C-indexes of 0.917 (training), 0.903 (internal testing), 0.942 (external testing set 1), and 0.880 (external testing set 2). The inclusion of mediastinal fat significantly improved predictive performance (NRI = 0.243, p < 0.05). A decision curve analysis indicated that incorporating mediastinal fat features provided greater patient benefits. Mediastinal fat offered complementary information for distinguishing benign from malignant nodules, enhancing the diagnostic capability of this deep learning-based radiomics model. This model demonstrated strong diagnostic ability for benign and malignant pulmonary nodules, providing a more accurate and beneficial approach for patient care.

Habitat Imaging With Tumoral and Peritumoral Radiomics for Prediction of Lung Adenocarcinoma Invasiveness on Preoperative Chest CT: A Multicenter Study.

Background: Tumors' growth processes result in spatial heterogeneity, with the development of tumor subregions (i.e., habitats) having unique biologic characteristics. Objective: To develop and validate a habitat model combining tumor and peritumoral radiomics features on chest CT for predicting invasiveness of lung adenocarcinoma. Methods: This retrospective study included 1156 patients (mean age, 57.5 years; 464 male, 692 female) from three centers and a public dataset, who underwent chest CT before lung adenocarcinoma resection (variable date ranges across datasets). Patients from one center formed training (n=500) and validation (n=215) sets; patients from the other sources formed three external test sets (n=249, 113, 79). For each patient, a single nodule was manually segmented on chest CT. The nodule segmentation was combined with an automatically generated 4-mm peritumoral region into a whole-volume volume-of-interest (VOI). A Gaussian mixture model (GMM) identified voxel clusters with similar first-order energy across patients. GMM results were used to divide each patient's whole-volume VOI into multiple habitats, defined consistently across patients. Radiomic features were extracted from each habitat. After feature selection, a habitat model was developed for predicting invasiveness, using pathologic assessment as a reference. An integrated model was constructed, combining features extracted from habitats and whole-volume VOIs. Model performance was evaluated, including in subgroups based on nodule density (pure ground-glass, part-solid, solid). Results: Invasive cancer was diagnosed in 625/1156 patients. GMM identified four as the optimal number of voxel clusters and thus of per-patient tumor habitats. The habitat model had AUC of 0.932 in the validation set, and 0.881, 0.880, and 0.764 in the three external test sets. The integrated model had AUC of 0.947 in the validation set and 0.936, 0.908, and 0.800 in the three external test sets. In the three external test sets combined, across nodule densities, AUCs for the habitat model were 0.836-0.969 and for the integrated model were 0.846-0.917. Conclusions: Habitat imaging combining tumoral and peritumoral radiomic features could help predict lung adenocarcinoma invasiveness. Prediction is improved when combining information on tumor subregions and the tumor overall. Clinical Impact: The findings may aid personalized preoperative assessments to guide clinical decision-making in lung adenocarcinoma.

The Impact of Race, Ethnicity, and Sex on Fairness in Artificial Intelligence for Glaucoma Prediction Models

ObjectiveDespite advances in artificial intelligence (AI) in glaucoma prediction, most works lack multi-center focus and do not consider fairness concerning sex, race, or ethnicity. This study aims to examine the impact of these sensitive attributes on developing fair AI models that predict glaucoma progression to necessitating incisional glaucoma surgery. DesignDatabase study. Participants39,090 patients with glaucoma, as identified by International Classification of Disease codes from 7 academic eye centers participating in the Sight OUtcomes Research Collaborative (SOURCE). MethodsWe developed XGBoost models using three approaches: (1) excluding sensitive attributes as input features, (2) including them explicitly as input features, and (3) training separate models for each group. Model input features included demographic details, diagnosis codes, medications, and clinical information (intraocular pressure, visual acuity etc.), from electronic health records. The models were trained on patients from 5 sites (N=27,999) and evaluated on a held-out internal test set (N=3,499) and 2 external test sets consisting of N=1,550 and N=2,542 patients. Main Outcomes and MeasuresArea under the receiver operating characteristic curve (AUROC) and equalized odds on the test set and external sites. Results6,682 (17.1%) of 39,090 patients underwent glaucoma surgery with a mean age of 70.1 (std. 14.6) years, 54.5% female, 62.3% White, 22.1% Black and 4.7% Latinx/Hispanic. We found that not including the sensitive attributes led to better classification performance (AUROC: 0.77-0.82) but worsened fairness when evaluated on the internal test set. However, on external test sites, the opposite was true: including sensitive attributes resulted in better classification performance (AUROC: external #1 - (0.73-0.81), external #2 - (0.67-0.70)), but varying degrees of fairness for sex and race as measured by equalized odds. ConclusionsAI models predicting whether patients with glaucoma progress to surgery demonstrated bias with respect to sex, race, and ethnicity. The effect of sensitive attribute inclusion and exclusion on fairness and performance varied based on internal versus external test sets. Prior to deployment, AI models should be evaluated for fairness on the target population.

Improving Computer-aided Detection for Digital Breast Tomosynthesis by Incorporating Temporal Change.

Purpose To develop a deep learning algorithm that uses temporal information to improve the performance of a previously published framework of cancer lesion detection for digital breast tomosynthesis. Materials and Methods This retrospective study analyzed the current and the 1-year-prior Hologic digital breast tomosynthesis screening examinations from eight different institutions between 2016 and 2020. The dataset contained 973 cancer and 7123 noncancer cases. The front end of this algorithm was an existing deep learning framework that performed single-view lesion detection followed by ipsilateral view matching. For this study, PriorNet was implemented as a cascaded deep learning module that used the additional growth information to refine the final probability of malignancy. Data from seven of the eight sites were used for training and validation, while the eighth site was reserved for external testing. Model performance was evaluated using localization receiver operating characteristic curves. Results On the validation set, PriorNet showed an area under the receiver operating characteristic curve (AUC) of 0.931 (95% CI: 0.930, 0.931), which outperformed both baseline models using single-view detection (AUC, 0.892 [95% CI: 0.891, 0.892]; P < .001) and ipsilateral matching (AUC, 0.915 [95% CI: 0.914, 0.915]; P < .001). On the external test set, PriorNet achieved an AUC of 0.896 (95% CI: 0.885, 0.896), outperforming both baselines (AUC, 0.846 [95% CI: 0.846, 0.847]; P < .001 and AUC, 0.865 [95% CI: 0.865, 0.866]; P < .001, respectively). In the high sensitivity range of 0.9 to 1.0, the partial AUC of PriorNet was significantly higher (P < .001) relative to both baselines. Conclusion PriorNet using temporal information further improved the breast cancer detection performance of an existing digital breast tomosynthesis cancer detection framework. Keywords: Digital Breast Tomosynthesis, Computer-aided Detection, Breast Cancer, Deep Learning © RSNA, 2024 See also commentary by Lee in this issue.

Flow Injection Analysis With UV Detection Versus Raman Spectroscopy for the Quantitative Analysis of Remdesivir

ABSTRACTThe quality and the security of preparations in hospital centres are essential to guarantee the patient's safety. The development of fast and efficient analytical methods is required to control the finished product before use. In this context, the study aimed to compare the performance and interchangeability of two spectral analytical methods: the flow injection analysis (FIA) with UV detection and the Raman spectroscopy for the quality control of preparation before use to quantify the remdesivir as a SARS‐CoV‐2 drug. A quantitative study of remdesivir was performed using clinically relevant concentration solutions ranging from 0.25 to 1.625 mg.mL−1 in 0.9% NaCl. Samples were analysed by FIA‐UV at 245 nm and by a handheld Raman spectroscopy at 785 nm. Quantitative models were developed using a calibration set (n = 45 samples) and optimized using a validation set (n = 27). An external validation test set (n = 58) was used to compare the two methods by a Bland–Altman plot. Partial least square regression was used to analyse Raman spectra, while univariate analysis was performed at 245 nm for FIA‐UV. The regression coefficient was higher than 0.990 for both methods, and the root mean square error of prediction was 0.031 mg.mL−1 for Raman spectroscopy. The Bland–Altman plot confirmed the interchangeability of the two methods and the potential of Raman spectroscopy to control remdesivir during clinical preparation in the hospital.

A deep learning system for myopia onset prediction and intervention effectiveness evaluation in children

The increasing prevalence of myopia worldwide presents a significant public health challenge. A key strategy to combat myopia is with early detection and prediction in children as such examination allows for effective intervention using readily accessible imaging technique. To this end, we introduced DeepMyopia, an artificial intelligence (AI)-enabled decision support system to detect and predict myopia onset and facilitate targeted interventions for children at risk using routine retinal fundus images. Based on deep learning architecture, DeepMyopia had been trained and internally validated on a large cohort of retinal fundus images (n = 1,638,315) and then externally tested on datasets from seven sites in China (n = 22,060). Our results demonstrated robustness of DeepMyopia, with AUCs of 0.908, 0.813, and 0.810 for 1-, 2-, and 3-year myopia onset prediction with the internal test set, and AUCs of 0.796, 0.808, and 0.767 with the external test set. DeepMyopia also effectively stratified children into low- and high-risk groups (p < 0.001) in both test sets. In an emulated randomized controlled trial (eRCT) on the Shanghai outdoor cohort (n = 3303) where DeepMyopia showed effectiveness in myopia prevention compared to NonCyc-based model, with an adjusted relative reduction (ARR) of −17.8%, 95% CI: −29.4%, −6.4%. DeepMyopia-assisted interventions attained quality-adjusted life years (QALYs) of 0.75 (95% CI: 0.53, 1.04) per person and avoided blindness years of 13.54 (95% CI: 9.57, 18.83) per 1 million persons compared to natural lifestyle with no active intervention. Our findings demonstrated DeepMyopia as a reliable and efficient AI-based decision support system for intervention guidance for children.

Performance and Clinical Impact of Radiomics and 3D-CNN Models for the Diagnosis of Neurodegenerative Parkinsonian Syndromes on 18 F-FDOPA PET.

The aim of this study was to compare the performance and added clinical value of a semiautomated radiomics model and an automated 3-dimensinal convolutional neural network (3D-CNN) model for diagnosing neurodegenerative parkinsonian syndromes on 18 F-FDOPA PET images. This 2-center retrospective study included 687 patients with motor symptoms consistent with parkinsonian syndrome. All patients underwent 18 F-FDOPA brain PET scans, acquired on 3 PET systems from 2 different hospitals, and classified as pathological or nonpathological (by an expert nuclear physician). Artificial intelligence models were trained to replicate this medical expert's classification using 2 pipelines. The radiomics pipeline was semiautomated and involved manually segmenting the bilateral caudate and putamen nuclei; 43 radiomic features were extracted and combined using the support vector machine method. The deep learning pipeline was fully automatic and used a 3D-CNN model. Both models were trained on 417 patients and tested on an internal (n = 100) and an external (n = 170) test set. The final models' performance was evaluated using balanced accuracy and compared with that of a junior medical expert and nonexpert nuclear physician. On the internal test set, the 3D-CNN model outperformed the radiomic model with a balanced accuracy of 99% (vs 96%). It led to diagnostic performance similar to that of a junior medical expert (only 1 in 100 patients misclassified by both). On the external test set from a less experienced hospital, the 3D-CNN model allowed physicians to correctly reclassify the diagnosis of 10 out 170 patients (6%). The developed 3D-CNN model can automatically diagnose neurodegenerative parkinsonian syndromes, also reducing diagnostic errors by 6% in less-experienced hospitals.

Fully Automated Deep Learning Model to Detect Clinically Significant Prostate Cancer at MRI.

Background Multiparametric MRI can help identify clinically significant prostate cancer (csPCa) (Gleason score ≥7) but is limited by reader experience and interobserver variability. In contrast, deep learning (DL) produces deterministic outputs. Purpose To develop a DL model to predict the presence of csPCa by using patient-level labels without information about tumor location and to compare its performance with that of radiologists. Materials and Methods Data from patients without known csPCa who underwent MRI from January 2017 to December 2019 at one of multiple sites of a single academic institution were retrospectively reviewed. A convolutional neural network was trained to predict csPCa from T2-weighted images, diffusion-weighted images, apparent diffusion coefficient maps, and T1-weighted contrast-enhanced images. The reference standard was pathologic diagnosis. Radiologist performance was evaluated as follows: Radiology reports were used for the internal test set, and four radiologists' PI-RADS ratings were used for the external (ProstateX) test set. The performance was compared using areas under the receiver operating characteristic curves (AUCs) and the DeLong test. Gradient-weighted class activation maps (Grad-CAMs) were used to show tumor localization. Results Among 5735 examinations in 5215 patients (mean age, 66 years ± 8 [SD]; all male), 1514 examinations (1454 patients) showed csPCa. In the internal test set (400 examinations), the AUC was 0.89 and 0.89 for the DL classifier and radiologists, respectively (P = .88). In the external test set (204 examinations), the AUC was 0.86 and 0.84 for the DL classifier and radiologists, respectively (P = .68). DL classifier plus radiologists had an AUC of 0.89 (P < .001). Grad-CAMs demonstrated activation over the csPCa lesion in 35 of 38 and 56 of 58 true-positive examinations in internal and external test sets, respectively. Conclusion The performance of a DL model was not different from that of radiologists in the detection of csPCa at MRI, and Grad-CAMs localized the tumor. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Johnson and Chandarana in this issue.