RESUMO
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.
Assuntos
Aprendizado Profundo , Imageamento por Ressonância Magnética , Neoplasias da Próstata , Masculino , Humanos , Neoplasias da Próstata/diagnóstico por imagem , Estudos Retrospectivos , Idoso , Pessoa de Meia-Idade , Imageamento por Ressonância Magnética/métodos , Interpretação de Imagem Assistida por Computador/métodos , Imageamento por Ressonância Magnética Multiparamétrica/métodos , Próstata/diagnóstico por imagem , Próstata/patologiaRESUMO
BACKGROUND: This study introduces THA-Net, a deep learning inpainting algorithm for simulating postoperative total hip arthroplasty (THA) radiographs from a single preoperative pelvis radiograph input, while being able to generate predictions either unconditionally (algorithm chooses implants) or conditionally (surgeon chooses implants). METHODS: The THA-Net is a deep learning algorithm which receives an input preoperative radiograph and subsequently replaces the target hip joint with THA implants to generate a synthetic yet realistic postoperative radiograph. We trained THA-Net on 356,305 pairs of radiographs from 14,357 patients from a single institution's total joint registry and evaluated the validity (quality of surgical execution) and realism (ability to differentiate real and synthetic radiographs) of its outputs against both human-based and software-based criteria. RESULTS: The surgical validity of synthetic postoperative radiographs was significantly higher than their real counterparts (mean difference: 0.8 to 1.1 points on 10-point Likert scale, P < .001), but they were not able to be differentiated in terms of realism in blinded expert review. Synthetic images showed excellent validity and realism when analyzed with already validated deep learning models. CONCLUSION: We developed a THA next-generation templating tool that can generate synthetic radiographs graded higher on ultimate surgical execution than real radiographs from training data. Further refinement of this tool may potentiate patient-specific surgical planning and enable technologies such as robotics, navigation, and augmented reality (an online demo of THA-Net is available at: https://demo.osail.ai/tha_net).
Assuntos
Artroplastia de Quadril , Aprendizado Profundo , Prótese de Quadril , Humanos , Artroplastia de Quadril/métodos , Articulação do Quadril/diagnóstico por imagem , Articulação do Quadril/cirurgia , Radiografia , Estudos RetrospectivosRESUMO
BACKGROUND: Revision total hip arthroplasty (THA) requires preoperatively identifying in situ implants, a time-consuming and sometimes unachievable task. Although deep learning (DL) tools have been attempted to automate this process, existing approaches are limited by classifying few femoral and zero acetabular components, only classify on anterior-posterior (AP) radiographs, and do not report prediction uncertainty or flag outlier data. METHODS: This study introduces Total Hip Arhtroplasty Automated Implant Detector (THA-AID), a DL tool trained on 241,419 radiographs that identifies common designs of 20 femoral and 8 acetabular components from AP, lateral, or oblique views and reports prediction uncertainty using conformal prediction and outlier detection using a custom framework. We evaluated THA-AID using internal, external, and out-of-domain test sets and compared its performance with human experts. RESULTS: THA-AID achieved internal test set accuracies of 98.9% for both femoral and acetabular components with no significant differences based on radiographic view. The femoral classifier also achieved 97.0% accuracy on the external test set. Adding conformal prediction increased true label prediction by 0.1% for acetabular and 0.7 to 0.9% for femoral components. More than 99% of out-of-domain and >89% of in-domain outlier data were correctly identified by THA-AID. CONCLUSIONS: The THA-AID is an automated tool for implant identification from radiographs with exceptional performance on internal and external test sets and no decrement in performance based on radiographic view. Importantly, this is the first study in orthopedics to our knowledge including uncertainty quantification and outlier detection of a DL model.
Assuntos
Artroplastia de Quadril , Aprendizado Profundo , Prótese de Quadril , Humanos , Incerteza , Acetábulo/cirurgia , Estudos RetrospectivosRESUMO
In recent years, deep learning (DL) has shown impressive performance in radiologic image analysis. However, for a DL model to be useful in a real-world setting, its confidence in a prediction must also be known. Each DL model's output has an estimated probability, and these estimated probabilities are not always reliable. Uncertainty represents the trustworthiness (validity) of estimated probabilities. The higher the uncertainty, the lower the validity. Uncertainty quantification (UQ) methods determine the uncertainty level of each prediction. Predictions made without UQ methods are generally not trustworthy. By implementing UQ in medical DL models, users can be alerted when a model does not have enough information to make a confident decision. Consequently, a medical expert could reevaluate the uncertain cases, which would eventually lead to gaining more trust when using a model. This review focuses on recent trends using UQ methods in DL radiologic image analysis within a conceptual framework. Also discussed in this review are potential applications, challenges, and future directions of UQ in DL radiologic image analysis.
Assuntos
Aprendizado Profundo , Radiologia , Humanos , Incerteza , Processamento de Imagem Assistida por ComputadorRESUMO
BACKGROUND: Whole-body low-dose CT is the recommended initial imaging modality to evaluate bone destruction as a result of multiple myeloma. Accurate interpretation of these scans to detect small lytic bone lesions is time intensive. A functional deep learning) algorithm to detect lytic lesions on CTs could improve the value of these CTs for myeloma imaging. Our objectives were to develop a DL algorithm and determine its performance at detecting lytic lesions of multiple myeloma. METHODS: Axial slices (2-mm section thickness) from whole-body low-dose CT scans of subjects with biochemically confirmed plasma cell dyscrasias were included in the study. Data were split into train and test sets at the patient level targeting a 90%/10% split. Two musculoskeletal radiologists annotated lytic lesions on the images with bounding boxes. Subsequently, we developed a two-step deep learning model comprising bone segmentation followed by lesion detection. Unet and "You Look Only Once" (YOLO) models were used as bone segmentation and lesion detection algorithms, respectively. Diagnostic performance was determined using the area under the receiver operating characteristic curve (AUROC). RESULTS: Forty whole-body low-dose CTs from 40 subjects yielded 2193 image slices. A total of 5640 lytic lesions were annotated. The two-step model achieved a sensitivity of 91.6% and a specificity of 84.6%. Lesion detection AUROC was 90.4%. CONCLUSION: We developed a deep learning model that detects lytic bone lesions of multiple myeloma on whole-body low-dose CTs with high performance. External validation is required prior to widespread adoption in clinical practice.
Assuntos
Aprendizado Profundo , Mieloma Múltiplo , Osteólise , Humanos , Mieloma Múltiplo/diagnóstico por imagem , Mieloma Múltiplo/patologia , Algoritmos , Tomografia Computadorizada por Raios X/métodosRESUMO
BACKGROUND: Defining standards is the first step toward quality assurance and improvement of educational programs. This study aimed at developing and validating a set of national standards for the Undergraduate Medical Education (UME) program through an accreditation system in Iran using the World Federation for Medical Education (WFME) framework. METHODS: The first draft of standards was prepared through consultative workshops with the participation of different UME program stakeholders. Subsequently, standards were sent to medical schools and UME directors were asked to complete a web-based survey. The content validity index at the item level (I-CVI) was computed using criteria including clarity, relevance, optimization and evaluability for each standard. Afterward, a full-day consultative workshop was held and a wide range of UME stakeholders across the country (n = 150) discussed the survey results and made corrections to standards. RESULTS: Analysis of survey results showed that relevance criteria had the best CVI as only 15 (13%) standards demonstrated CVI < 0.78. More than two-thirds (71%) and a half (55%) of standards showed CVI < 0.78 for optimization and evaluability criteria. The final set of UME national standards was structured in 9 areas, 24 sub-areas, 82 basic and 40 quality development standards, and 84 annotations. CONCLUSIONS: We developed and validated national standards as a framework to ensure the quality of UME training with input from UME stakeholders. We used WFME standards as a benchmark while addressing local requirements. The standards and participatory approach to developing standards may guide relevant institutions.
Assuntos
Educação de Graduação em Medicina , Educação Médica , Humanos , Irã (Geográfico) , Acreditação , BenchmarkingRESUMO
BACKGROUND: Automatic methods for labeling and segmenting pelvis structures can improve the efficiency of clinical and research workflows and reduce the variability introduced with manual labeling. The purpose of this study was to develop a single deep learning model to annotate certain anatomical structures and landmarks on antero-posterior (AP) pelvis radiographs. METHODS: A total of 1,100 AP pelvis radiographs were manually annotated by 3 reviewers. These images included a mix of preoperative and postoperative images as well as a mix of AP pelvis and hip images. A convolutional neural network was trained to segment 22 different structures (7 points, 6 lines, and 9 shapes). Dice score, which measures overlap between model output and ground truth, was calculated for the shapes and lines structures. Euclidean distance error was calculated for point structures. RESULTS: Dice score averaged across all images in the test set was 0.88 and 0.80 for the shape and line structures, respectively. For the 7-point structures, average distance between real and automated annotations ranged from 1.9 mm to 5.6 mm, with all averages falling below 3.1 mm except for the structure labeling the center of the sacrococcygeal junction, where performance was low for both human and machine-produced labels. Blinded qualitative evaluation of human and machine produced segmentations did not reveal any drastic decrease in performance of the automatic method. CONCLUSION: We present a deep learning model for automated annotation of pelvis radiographs that flexibly handles a variety of views, contrasts, and operative statuses for 22 structures and landmarks.
Assuntos
Aprendizado Profundo , Humanos , Radiografia , Redes Neurais de Computação , Pelve/diagnóstico por imagem , Período Pós-OperatórioRESUMO
BACKGROUND: In this work, we applied and validated an artificial intelligence technique known as generative adversarial networks (GANs) to create large volumes of high-fidelity synthetic anteroposterior (AP) pelvis radiographs that can enable deep learning (DL)-based image analyses, while ensuring patient privacy. METHODS: AP pelvis radiographs with native hips were gathered from an institutional registry between 1998 and 2018. The data was used to train a model to create 512 × 512 pixel synthetic AP pelvis images. The network was trained on 25 million images produced through augmentation. A set of 100 random images (50/50 real/synthetic) was evaluated by 3 orthopaedic surgeons and 2 radiologists to discern real versus synthetic images. Two models (joint localization and segmentation) were trained using synthetic images and tested on real images. RESULTS: The final model was trained on 37,640 real radiographs (16,782 patients). In a computer assessment of image fidelity, the final model achieved an "excellent" rating. In a blinded review of paired images (1 real, 1 synthetic), orthopaedic surgeon reviewers were unable to correctly identify which image was synthetic (accuracy = 55%, Kappa = 0.11), highlighting synthetic image fidelity. The synthetic and real images showed equivalent performance when they were assessed by established DL models. CONCLUSION: This work shows the ability to use a DL technique to generate a large volume of high-fidelity synthetic pelvis images not discernible from real imaging by computers or experts. These images can be used for cross-institutional sharing and model pretraining, further advancing the performance of DL models without risk to patient data safety. LEVEL OF EVIDENCE: Level III.
Assuntos
Aprendizado Profundo , Humanos , Inteligência Artificial , Privacidade , Processamento de Imagem Assistida por Computador/métodos , Pelve/diagnóstico por imagemRESUMO
BACKGROUND: Many risk factors have been described for periprosthetic femur fracture (PPFFx) following total hip arthroplasty (THA), yet a patient-specific risk assessment tool remains elusive. The purpose of this study was to develop a high-dimensional, patient-specific risk-stratification nomogram that allows dynamic risk modification based on operative decisions. METHODS: We evaluated 16,696 primary nononcologic THAs performed between 1998 and 2018. During a mean 6-year follow-up, 558 patients (3.3%) sustained a PPFFx. Patients were characterized by individual natural language processing-assisted chart review on nonmodifiable factors (demographics, THA indication, and comorbidities), and modifiable operative decisions (femoral fixation [cemented/uncemented], surgical approach [direct anterior, lateral, and posterior], and implant type [collared/collarless]). Multivariable Cox regression models and nomograms were developed with PPFFx as a binary outcome at 90 days, 1 year, and 5 years, postoperatively. RESULTS: Patient-specific PPFFx risk based on comorbid profile was wide-ranging from 0.4-18% at 90 days, 0.4%-20% at 1 year, and 0.5%-25% at 5 years. Among 18 evaluated patient factors, 7 were retained in multivariable analyses. The 4 significant nonmodifiable factors included the following: women (hazard ratio (HR) = 1.6), older age (HR = 1.2 per 10 years), diagnosis of osteoporosis or use of osteoporosis medications (HR = 1.7), and indication for surgery other than osteoarthritis (HR = 2.2 for fracture, HR = 1.8 for inflammatory arthritis, HR = 1.7 for osteonecrosis). The 3 modifiable surgical factors were included as follows: uncemented femoral fixation (HR = 2.5), collarless femoral implants (HR = 1.3), and surgical approach other than direct anterior (lateral HR = 2.9, posterior HR = 1.9). CONCLUSION: This patient-specific PPFFx risk calculator demonstrated a wide-ranging risk based on comorbid profile and enables surgeons to quantify risk mitigation based on operative decisions. LEVEL OF EVIDENCE: Level III, Prognostic.
Assuntos
Artroplastia de Quadril , Distinções e Prêmios , Fraturas do Fêmur , Prótese de Quadril , Fraturas Periprotéticas , Humanos , Feminino , Artroplastia de Quadril/efeitos adversos , Artroplastia de Quadril/métodos , Fraturas Periprotéticas/epidemiologia , Fraturas Periprotéticas/etiologia , Fraturas Periprotéticas/cirurgia , Prótese de Quadril/efeitos adversos , Reoperação , Fraturas do Fêmur/epidemiologia , Fraturas do Fêmur/etiologia , Fraturas do Fêmur/cirurgia , Fatores de Risco , Estudos RetrospectivosRESUMO
Since 2000, there have been more than 8000 publications on radiology artificial intelligence (AI). AI breakthroughs allow complex tasks to be automated and even performed beyond human capabilities. However, the lack of details on the methods and algorithm code undercuts its scientific value. Many science subfields have recently faced a reproducibility crisis, eroding trust in processes and results, and influencing the rise in retractions of scientific papers. For the same reasons, conducting research in deep learning (DL) also requires reproducibility. Although several valuable manuscript checklists for AI in medical imaging exist, they are not focused specifically on reproducibility. In this study, we conducted a systematic review of recently published papers in the field of DL to evaluate if the description of their methodology could allow the reproducibility of their findings. We focused on the Journal of Digital Imaging (JDI), a specialized journal that publishes papers on AI and medical imaging. We used the keyword "Deep Learning" and collected the articles published between January 2020 and January 2022. We screened all the articles and included the ones which reported the development of a DL tool in medical imaging. We extracted the reported details about the dataset, data handling steps, data splitting, model details, and performance metrics of each included article. We found 148 articles. Eighty were included after screening for articles that reported developing a DL model for medical image analysis. Five studies have made their code publicly available, and 35 studies have utilized publicly available datasets. We provided figures to show the ratio and absolute count of reported items from included studies. According to our cross-sectional study, in JDI publications on DL in medical imaging, authors infrequently report the key elements of their study to make it reproducible.
Assuntos
Inteligência Artificial , Diagnóstico por Imagem , Humanos , Estudos Transversais , Reprodutibilidade dos Testes , AlgoritmosRESUMO
BACKGROUND AND AIMS: The risk of progression in Barrett's esophagus (BE) increases with development of dysplasia. There is a critical need to improve the diagnosis of BE dysplasia, given substantial interobserver disagreement among expert pathologists and overdiagnosis of dysplasia by community pathologists. We developed a deep learning model to predict dysplasia grade on whole-slide imaging. METHODS: We digitized nondysplastic BE (NDBE), low-grade dysplasia (LGD), and high-grade dysplasia (HGD) histology slides. Two expert pathologists confirmed all histology and digitally annotated areas of dysplasia. Training, validation, and test sets were created (by a random 70/20/10 split). We used an ensemble approach combining a "you only look once" model to identify regions of interest and histology class (NDBE, LGD, or HGD) followed by a ResNet101 model pretrained on ImageNet applied to the regions of interest. Diagnostic performance was determined for the whole slide. RESULTS: We included slides from 542 patients (164 NDBE, 226 LGD, and 152 HGD) yielding 8596 bounding boxes in the training set, 1946 bounding boxes in the validation set, and 840 boxes in the test set. When the ensemble model was used, sensitivity and specificity for LGD was 81.3% and 100%, respectively, and >90% for NDBE and HGD. The overall positive predictive value and sensitivity metric (calculated as F1 score) was .91 for NDBE, .90 for LGD, and 1.0 for HGD. CONCLUSIONS: We successfully trained and validated a deep learning model to accurately identify dysplasia on whole-slide images. This model can potentially help improve the histologic diagnosis of BE dysplasia and the appropriate application of endoscopic therapy.
Assuntos
Adenocarcinoma , Esôfago de Barrett , Aprendizado Profundo , Neoplasias Esofágicas , Humanos , Esôfago de Barrett/diagnóstico , Esôfago de Barrett/patologia , Neoplasias Esofágicas/patologia , Adenocarcinoma/patologia , Progressão da Doença , HiperplasiaRESUMO
BACKGROUND: Several studies have investigated the distribution of hip-knee-ankle (HKA) angle in healthy populations; however, few have evaluated this metric in patients undergoing total knee arthroplasty (TKA). The purpose of this study is to compare HKA angle distribution in early and advanced knee osteoarthritis (OA) patients. METHODS: Full limb radiographs were used to measure HKA angle for 983 subjects from the Osteoarthritis Initiative (OAI) cohort and 4,901 pre-TKA patients from an institutional cohort. Measurements were made using a previously validated deep learning algorithm. Linear regression models were used to determine the association of HKA alignment angle with patient characteristics. RESULTS: The mean ± standard deviation HKA angle was -1.3° ± 3.2° in the OAI cohort and -4.1° ± 6.1° in the pre-TKA cohort. In the OAI cohort, normal alignment (64%) was the most common knee alignment followed by varus (29%), and valgus (7%). In pre-TKA patients, the most common alignment was varus (62%), followed by normal (27%) and valgus (11%). In pre-TKA patients, mean HKA angle in primary knee OA, post-traumatic knee OA, and rheumatoid arthritis patients were -4.3° ± 6.1°, -3.2° ± 6.4°, and -2.9° ± 6.1°, respectively. HKA angle was strongly associated (P < .001) with gender and body mass index. CONCLUSION: TKA patients have a wider alignment distribution and more severe varus and valgus alignment than individuals "at risk" for knee OA from the OAI cohort. These epidemiologic findings improve our understanding of HKA angle distribution and its correlation with demographic characteristics in early and late-stage arthritis.
Assuntos
Artroplastia do Joelho , Osteoartrite do Joelho , Tornozelo/cirurgia , Articulação do Tornozelo/cirurgia , Humanos , Articulação do Joelho/diagnóstico por imagem , Articulação do Joelho/cirurgia , Osteoartrite do Joelho/cirurgia , Estudos RetrospectivosRESUMO
BACKGROUND: Dislocation is a common complication following total hip arthroplasty (THA), and accounts for a high percentage of subsequent revisions. The purpose of this study is to illustrate the potential of a convolutional neural network model to assess the risk of hip dislocation based on postoperative anteroposterior pelvis radiographs. METHODS: We retrospectively evaluated radiographs for a cohort of 13,970 primary THAs with 374 dislocations over 5 years of follow-up. Overall, 1490 radiographs from dislocated and 91,094 from non-dislocated THAs were included in the analysis. A convolutional neural network object detection model (YOLO-V3) was trained to crop the images by centering on the femoral head. A ResNet18 classifier was trained to predict subsequent hip dislocation from the cropped imaging. The ResNet18 classifier was initialized with ImageNet weights and trained using FastAI (V1.0) running on PyTorch. The training was run for 15 epochs using 10-fold cross validation, data oversampling, and augmentation. RESULTS: The hip dislocation classifier achieved the following mean performance (standard deviation): accuracy = 49.5 (4.1%), sensitivity = 89.0 (2.2%), specificity = 48.8 (4.2%), positive predictive value = 3.3 (0.3%), negative predictive value = 99.5 (0.1%), and area under the receiver operating characteristic curve = 76.7 (3.6%). Saliency maps demonstrated that the model placed the greatest emphasis on the femoral head and acetabular component. CONCLUSION: Existing prediction methods fail to identify patients at high risk of dislocation following THA. Our radiographic classifier model has high sensitivity and negative predictive value, and can be combined with clinical risk factor information for rapid assessment of risk for dislocation following THA. The model further suggests radiographic locations which may be important in understanding the etiology of prosthesis dislocation. Importantly, our model is an illustration of the potential of automated imaging artificial intelligence models in orthopedics. LEVEL OF EVIDENCE: Level III.
Assuntos
Artroplastia de Quadril , Aprendizado Profundo , Luxação do Quadril , Prótese de Quadril , Artroplastia de Quadril/efeitos adversos , Inteligência Artificial , Luxação do Quadril/diagnóstico por imagem , Luxação do Quadril/epidemiologia , Prótese de Quadril/efeitos adversos , Humanos , Estudos Retrospectivos , Fatores de RiscoRESUMO
BACKGROUND: Inappropriate acetabular component angular position is believed to increase the risk of hip dislocation after total hip arthroplasty. However, manual measurement of these angles is time consuming and prone to interobserver variability. The purpose of this study was to develop a deep learning tool to automate the measurement of acetabular component angles on postoperative radiographs. METHODS: Two cohorts of 600 anteroposterior (AP) pelvis and 600 cross-table lateral hip postoperative radiographs were used to develop deep learning models to segment the acetabular component and the ischial tuberosities. Cohorts were manually annotated, augmented, and randomly split to train-validation-test data sets on an 8:1:1 basis. Two U-Net convolutional neural network models (one for AP and one for cross-table lateral radiographs) were trained for 50 epochs. Image processing was then deployed to measure the acetabular component angles on the predicted masks for anatomical landmarks. Performance of the tool was tested on 80 AP and 80 cross-table lateral radiographs. RESULTS: The convolutional neural network models achieved a mean Dice similarity coefficient of 0.878 and 0.903 on AP and cross-table lateral test data sets, respectively. The mean difference between human-level and machine-level measurements was 1.35° (σ = 1.07°) and 1.39° (σ = 1.27°) for the inclination and anteversion angles, respectively. Differences of 5° or more between human-level and machine-level measurements were observed in less than 2.5% of cases. CONCLUSION: We developed a highly accurate deep learning tool to automate the measurement of angular position of acetabular components for use in both clinical and research settings. LEVEL OF EVIDENCE: III.
Assuntos
Artroplastia de Quadril , Aprendizado Profundo , Prótese de Quadril , Acetábulo/diagnóstico por imagem , Acetábulo/cirurgia , Artroplastia de Quadril/efeitos adversos , Prótese de Quadril/efeitos adversos , Humanos , RadiografiaRESUMO
Thyroid Ultrasound (US) is the primary method to evaluate thyroid nodules. Deep learning (DL) has been playing a significant role in evaluating thyroid cancer. We propose a DL-based pipeline to detect and classify thyroid nodules into benign or malignant groups relying on two views of US imaging. Transverse and longitudinal US images of thyroid nodules from 983 patients were collected retrospectively. Eighty-one cases were held out as a testing set, and the rest of the data were used in five-fold cross-validation (CV). Two You Look Only Once (YOLO) v5 models were trained to detect nodules and classify them. For each view, five models were developed during the CV, which was ensembled by using non-max suppression (NMS) to boost their collective generalizability. An extreme gradient boosting (XGBoost) model was trained on the outputs of the ensembled models for both views to yield a final prediction of malignancy for each nodule. The test set was evaluated by an expert radiologist using the American College of Radiology Thyroid Imaging Reporting and Data System (ACR-TIRADS). The ensemble models for each view achieved a mAP0.5 of 0.797 (transverse) and 0.716 (longitudinal). The whole pipeline reached an AUROC of 0.84 (CI 95%: 0.75-0.91) with sensitivity and specificity of 84% and 63%, respectively, while the ACR-TIRADS evaluation of the same set had a sensitivity of 76% and specificity of 34% (p-value = 0.003). Our proposed work demonstrated the potential possibility of a deep learning model to achieve diagnostic performance for thyroid nodule evaluation.
RESUMO
In recent years, the role of Artificial Intelligence (AI) in medical imaging has become increasingly prominent, with the majority of AI applications approved by the FDA being in imaging and radiology in 2023. The surge in AI model development to tackle clinical challenges underscores the necessity for preparing high-quality medical imaging data. Proper data preparation is crucial as it fosters the creation of standardized and reproducible AI models while minimizing biases. Data curation transforms raw data into a valuable, organized, and dependable resource and is a fundamental process to the success of machine learning and analytical projects. Considering the plethora of available tools for data curation in different stages, it is crucial to stay informed about the most relevant tools within specific research areas. In the current work, we propose a descriptive outline for different steps of data curation while we furnish compilations of tools collected from a survey applied among members of the Society of Imaging Informatics (SIIM) for each of these stages. This collection has the potential to enhance the decision-making process for researchers as they select the most appropriate tool for their specific tasks.