Generalizability and Diagnostic Performance of AI Models for Thyroid US.

Background Artificial intelligence (AI) models have improved US assessment of thyroid nodules; however, the lack of generalizability limits the application of these models. Purpose To develop AI models for segmentation and classification of thyroid nodules in US using diverse data sets from nationwide hospitals and multiple vendors, and to measure the impact of the AI models on diagnostic performance. Materials and Methods This retrospective study included consecutive patients with pathologically confirmed thyroid nodules who underwent US using equipment from 12 vendors at 208 hospitals across China from November 2017 to January 2019. The detection, segmentation, and classification models were developed based on the subset or complete set of images. Model performance was evaluated by precision and recall, Dice coefficient, and area under the receiver operating characteristic curve (AUC) analyses. Three scenarios (diagnosis without AI assistance, with freestyle AI assistance, and with rule-based AI assistance) were compared with three senior and three junior radiologists to optimize incorporation of AI into clinical practice. Results A total of 10 023 patients (median age, 46 years [IQR 37-55 years]; 7669 female) were included. The detection, segmentation, and classification models had an average precision, Dice coefficient, and AUC of 0.98 (95% CI: 0.96, 0.99), 0.86 (95% CI: 0.86, 0.87), and 0.90 (95% CI: 0.88, 0.92), respectively. The segmentation model trained on the nationwide data and classification model trained on the mixed vendor data exhibited the best performance, with a Dice coefficient of 0.91 (95% CI: 0.90, 0.91) and AUC of 0.98 (95% CI: 0.97, 1.00), respectively. The AI model outperformed all senior and junior radiologists (P < .05 for all comparisons), and the diagnostic accuracies of all radiologists were improved (P < .05 for all comparisons) with rule-based AI assistance. Conclusion Thyroid US AI models developed from diverse data sets had high diagnostic performance among the Chinese population. Rule-based AI assistance improved the performance of radiologists in thyroid cancer diagnosis. © RSNA, 2023 Supplemental material is available for this article.

Development of machine learning model to predict pulmonary function with low-dose CT-derived parameter response mapping in a community-based chest screening cohort.

PURPOSE: To construct and evaluate the performance of a machine learning-based low dose computed tomography (LDCT)-derived parametric response mapping (PRM) model for predicting pulmonary function test (PFT) results. MATERIALS AND METHODS: A total of 615 subjects from a community-based screening population (40-74 years old) with PFT parameters, including the ratio of the first second forced expiratory volume to forced vital capacity (FEV1/FVC), the percentage of forced expiratory volume in the one second predicted (FEV1%), and registered inspiration-to-expiration chest CT scanning were enrolled retrospectively. Subjects were classified into a normal, high risk, and COPD group based on PFT. Data of 72 PRM-derived quantitative parameters were collected, including volume and volume percentage of emphysema, functional-small airways disease, and normal lung tissue. A machine-learning with random forest regression model and a multilayer perceptron (MLP) model were constructed and tested on PFT prediction, which was followed by evaluation of classification performance based on the PFT predictions. RESULTS: The machine-learning model based on PRM parameters showed better performance for predicting PFT than MLP, with a coefficient of determination (R2 ) of 0.749 and 0.792 for FEV1/FVC and FEV1%, respectively. The Mean Squared Errors (MSE) for FEV1/FVC and FEV1% are 0.0030 and 0.0097 for the random forest model, respectively. The Root Mean Squared Errors (RMSE) for FEV1/FVC and FEV1% are 0.055 and 0.098, respectively. The sensitivity, specificity, and accuracy for differentiating between the normal group and high-risk group were 34/40 (85%), 65/72 (90%), and 99/112 (88%), respectively. For differentiating between the non-COPD group and COPD group, the sensitivity, specificity, and accuracy were 8/9 (89%), 112/112 (100%), 120/121 (99%), respectively. CONCLUSIONS: The machine learning-based random forest model predicts PFT results in a community screening population based on PRM, and it identifies high risk COPD from normal populations with high sensitivity and reliably predicts of high-risk COPD.

Deep convolutional neural network-based classification of cancer cells on cytological pleural effusion images.

Lung cancer is one of the leading causes of cancer-related death worldwide. Cytology plays an important role in the initial evaluation and diagnosis of patients with lung cancer. However, due to the subjectivity of cytopathologists and the region-dependent diagnostic levels, the low consistency of liquid-based cytological diagnosis results in certain proportions of misdiagnoses and missed diagnoses. In this study, we performed a weakly supervised deep learning method for the classification of benign and malignant cells in lung cytological images through a deep convolutional neural network (DCNN). A total of 404 cases of lung cancer cells in effusion cytology specimens from Shanghai Pulmonary Hospital were investigated, in which 266, 78, and 60 cases were used as the training, validation and test sets, respectively. The proposed method was evaluated on 60 whole-slide images (WSIs) of lung cancer pleural effusion specimens. This study showed that the method had an accuracy, sensitivity, and specificity respectively of 91.67%, 87.50% and 94.44% in classifying malignant and benign lesions (or normal). The area under the receiver operating characteristic (ROC) curve (AUC) was 0.9526 (95% confidence interval (CI): 0.9019-9.9909). In contrast, the average accuracies of senior and junior cytopathologists were 98.34% and 83.34%, respectively. The proposed deep learning method will be useful and may assist pathologists with different levels of experience in the diagnosis of cancer cells on cytological pleural effusion images in the future.

Deep Learning Assisted Diagnosis of Musculoskeletal Tumors Based on Contrast-Enhanced Magnetic Resonance Imaging.

BACKGROUND: Misdiagnosis of malignant musculoskeletal tumors may lead to the delay of intervention, resulting in amputation or death. PURPOSE: To improve the diagnostic efficacy of musculoskeletal tumors by developing deep learning (DL) models based on contrast-enhanced magnetic resonance imaging and to quantify the improvement in diagnostic performance obtained by using these models. STUDY TYPE: Retrospective. POPULATION: Three hundreds and four musculoskeletal tumors, including 212 malignant and 92 benign lesions, were randomized into the training (n = 180), validation (n = 62) and testing cohort (n = 62). FIELD STRENGTH/SEQUENCE: A 3 T/T1 -weighted (T1 -w), T2 -weighted (T2 -w), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted (CET1 -w) images. ASSESSMENT: Three DL models based, respectively, on the sagittal, coronal, and axial MR images were constructed to predict the malignancy of tumors. Blinded to the prediction results, a group of specialists made independent initial diagnoses for each patient by reading all image sequences. One month after the initial diagnoses, the same group of doctors made another round of diagnoses knowing the malignancy of each tumor predicted by the three models. The reference standard was the pathological diagnosis of malignancy. STATISTICAL TESTS: Sensitivity, specificity, and accuracy (all with 95% confidential intervals [CI]) corresponding to each diagnostic test were computed. Chi-square tests were used to assess the differences in those parameters with and without DL models. A P value < 0.05 was considered statistically significant. RESULTS: The developed models significantly improved the diagnostic sensitivities of two oncologists by 0.15 (95% CI: 0.06-0.24) and 0.36 (95% CI: 0.24-0.28), one radiologist by 0.12 (95% CI: 0.04-0.20), and three of the four orthopedists, respectively, by 0.12 (95% CI: 0.04-0.20), 0.29 (95% CI: 0.18-0.40), and 0.23 (95% CI: 0.13-0.33), without impairing any of their diagnostic specificities (all P > 0.128). DATA CONCLUSION: The DL models developed can significantly improve the performance of doctors with different training and experience in diagnosing musculoskeletal tumors. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 2.

Aerodynamic Simulation of Small Airway Resistance: A New Imaging Biomarker for Chronic Obstructive Pulmonary Disease.

Purpose: To develop a novel method for calculating small airway resistance using computational fluid dynamics (CFD) based on CT data and evaluate its value to identify COPD. Patients and Methods: 24 subjects who underwent chest CT scans and pulmonary function tests between August 2020 and December 2020 were enrolled retrospectively. Subjects were divided into three groups: normal (10), high-risk (6), and COPD (8). The airway from the trachea down to the sixth generation of bronchioles was reconstructed by a 3D slicer. The small airway resistance (RSA) and RSA as a percentage of total airway resistance (RSA%) were calculated by CFD combined with airway resistance and FEV1 measured by pulmonary function test. A correlation analysis was conducted between RSA and pulmonary function parameters, including FEV1/FVC, FEV1% predicted, MEF50% predicted, MEF75% predicted and MMEF75/25% predicted. Results: The RSA and RSA% were significantly different among the three groups (p<0.05) and related to FEV1/FVC (r = -0.70, p < 0.001; r = -0.67, p < 0.001), FEV1% predicted (r = -0.60, p = 0.002; r = -0.57, p = 0.004), MEF50% predicted (r = -0.64, p = 0.001; r = -0.64, p = 0.001), MEF75% predicted (r = -0.71, p < 0.001; r = -0.60, p = 0.002) and MMEF 75/25% predicted (r = -0.64, p = 0.001; r = -0.64, p = 0.001). Conclusion: Airway CFD is a valuable method for estimating the small airway resistance, where the derived RSA will aid in the early diagnosis of COPD.

Unsupervised Local Discrimination for Medical Images.

Contrastive learning, which aims to capture general representation from unlabeled images to initialize the medical analysis models, has been proven effective in alleviating the high demand for expensive annotations. Current methods mainly focus on instance-wise comparisons to learn the global discriminative features, however, pretermitting the local details to distinguish tiny anatomical structures, lesions, and tissues. To address this challenge, in this paper, we propose a general unsupervised representation learning framework, named local discrimination (LD), to learn local discriminative features for medical images by closely embedding semantically similar pixels and identifying regions of similar structures across different images. Specifically, this model is equipped with an embedding module for pixel-wise embedding and a clustering module for generating segmentation. And these two modules are unified by optimizing our novel region discrimination loss function in a mutually beneficial mechanism, which enables our model to reflect structure information as well as measure pixel-wise and region-wise similarity. Furthermore, based on LD, we propose a center-sensitive one-shot landmark localization algorithm and a shape-guided cross-modality segmentation model to foster the generalizability of our model. When transferred to downstream tasks, the learned representation by our method shows a better generalization, outperforming representation from 18 state-of-the-art (SOTA) methods and winning 9 out of all 12 downstream tasks. Especially for the challenging lesion segmentation tasks, the proposed method achieves significantly better performance.

Radiologists with assistance of deep learning can achieve overall accuracy of benign-malignant differentiation of musculoskeletal tumors comparable with that of pre-surgical biopsies in the literature.

PURPOSE: The purpose of this study was to assess if radiologists assisted by deep learning (DL) algorithms can achieve diagnostic accuracy comparable to that of pre-surgical biopsies in benign-malignant differentiation of musculoskeletal tumors (MST). METHODS: We first conducted a systematic review of literature to get the respective overall diagnostic accuracies of fine-needle aspiration biopsy (FNAB) and core needle biopsy (CNB) in differentiating between benign and malignant MST, by synthesizing data from the articles meeting our inclusion criteria. To compared against the accuracies reported in literature, we then invited 4 radiologists, respectively with 2 (A), 6 (B), 7 (C), and 33 (D) years of experience in interpreting musculoskeletal MRI to perform diagnostic tests on our own dataset (n = 62), with and without assistance of a previously developed DL algorithm. The gold standard for benign-malignant differentiation was histopathologic confirmation or clinical/radiographic follow-up. RESULTS: For FNAB, a meta-analysis containing 4604 samples met the inclusion criteria, with the overall diagnostic accuracy reported to be 0.77. For CNB, an overall accuracy of 0.86 was derived by synthesizing results from 7 original research articles containing a total of 587 samples. On our internal MST dataset, the invited radiologists, respectively, achieved diagnostic accuracies of 0.84 (A), 0.89 (B), 0.87 (C), and 0.90 (D), with the assistance of DL. CONCLUSION: Use of DL algorithms on musculoskeletal dynamic contrast-enhanced MRI improved the benign-malignant differentiation accuracy of radiologists to a level comparable to that of pre-surgical biopsies. The developed DL algorithms have a potential to lower the risk of miss-diagnosing malignancy in radiological practice.

Deep learning to automatically evaluate HER2 gene amplification status from fluorescence in situ hybridization images.

Human epidermal growth factor receptor 2 (HER2) gene amplification helps identify breast cancer patients who may respond to targeted anti-HER2 therapy. This study aims to develop an automated method for quantifying HER2 fluorescence in situ hybridization (FISH) signals and improve the working efficiency of pathologists. An Aitrox artificial intelligence (AI) model based on deep learning was constructed, and a comparison between the AI model and traditional manual counting was performed. In total, 918 FISH images from 320 consecutive invasive breast cancers were analysed and automatically classified into 5 groups according to the 2018 ASCO/CAP guidelines. The overall classification accuracy was 85.33% (157/184) with a mean average precision of 0.735. In Group 5, the most common group, the consistency was as high as 95.90% (117/122), while the consistency was low in the other groups due to the limited number of cases. The causes of this inconsistency, including clustered HER2 signals, coarse CEP17 signals and some section quality problems, were analysed. The developed AI model is a reliable tool for evaluating HER2 amplification statuses, especially for breast cancer in Group 5; additional cases from multiple centres could further improve the accuracy achieved for other groups.

Molecular characterization of the bla(KPC-2) gene in clinical isolates of carbapenem-resistant Klebsiella pneumoniae from the pediatric wards of a Chinese hospital.

The present study was conducted to confirm the presence of Klebsiella pneumoniae carbapenemase (KPC)-producing K. pneumoniae associated with a nosocomial outbreak in a Chinese pediatric hospital. From July 2009 to January 2011, 124 nonduplicated K. pneumoniae isolates were collected from specimens from patients of pediatric units in the hospital. Twelve of the 124 isolates possessed the bla(KPC-2) gene and showed 7 different pulsed-field gel electrophoresis (PFGE) patterns. Meanwhile, 16S rRNA methylase, acc(6')-Ib-cr, and several types of ß-lactamases were also produced by the majority of the KPC-producing isolates. Class 1 integron-encoded intI1 integrase gene was subsequently found in all strains, and amplification, sequencing, and comparison of DNA between 5' conserved segment and 3' conserved segment region showed the presence of several known antibiotic resistance gene cassettes of various sizes. The conjugation and plasmid-curing experiments indicated some KPC-2-encoding genes were transmissible. In addition, conjugal cotransfer of multidrug-resistant phenotypes with KPC-positive phenotypes was observed in KPC-producing strains. Restriction endonuclease analysis and DNA hybridization with a KPC-specific probe showed that the bla(KPC-2) gene was carried by plasmid DNA from K. pneumoniae of PFGE pattern B. The overall results indicate that the emergence and outbreak of KPC-producing K. pneumoniae in our pediatric wards occurred in conjunction with plasmids coharboring 16S rRNA methylase and extended-spectrum ß-lactamases.

A new AI-assisted scoring system for PD-L1 expression in NSCLC.

BACKGROUND: Artificial intelligence (AI) analysis may serve as a scoring tool for programmed cell death ligand-1 (PD-L1) expression. In this study, a new AI-assisted scoring system for pathologists was tested for PD-L1 expression assessment in non-small cell lung cancer (NSCLC). METHODS: PD-L1 expression was evaluated using the tumor proportion score (TPS) categorized into three levels: negative (TPS < 1%), low expression (TPS 1-49%), and high expression (TPS ≥ 50%). In order to train, validate, and test the Aitrox AI segmentation model at the whole slide image (WSI) level, 54, 53, and 115 cases were used as training, validation, and test datasets, respectively. TPS reading results from five experienced pathologists, six inexperienced and the Aitrox AI model were analyzed on 115 PD-L1 stained WSIs. The Gold Standard for TPS was derived from the review of three expert pathologists. Spearman's correlation coefficient was calculated and compared between the results. RESULTS: Aitrox AI Model correlated strongly with the TPS Gold Standard and was comparable with the results of three of the five experienced pathologists. In contrast, the results of four of the six inexperienced pathologists correlated only moderately with the TPS Gold Standard. Aitrox AI Model performed better than the inexperienced pathologists and was comparable to experienced pathologists in both negative and low TPS groups. Despite the fact that the low TPS group showed 5.09% of cases with large fluctuations, the Aitrox AI Model still showed a higher correlation than the inexperienced pathologists. However, the AI model showed unsatisfactory performance in the high TPS groups, especially lower values than the Gold Standard in images with large regions of false-positive cells. CONCLUSION: The Aitrox AI Model demonstrates potential in assisting routine diagnosis of NSCLC by pathologists through scoring of PD-L1 expression.