RESUMO
Background: Most primary bone tumors are often found in the bone around the knee joint. However, the detection of primary bone tumors on radiographs can be challenging for the inexperienced or junior radiologist. This study aimed to develop a deep learning (DL) model for the detection of primary bone tumors around the knee joint on radiographs. Methods: From four tertiary referral centers, we recruited 687 patients diagnosed with bone tumors (including osteosarcoma, chondrosarcoma, giant cell tumor of bone, bone cyst, enchondroma, fibrous dysplasia, etc.; 417 males, 270 females; mean age 22.8±13.2 years) by postoperative pathology or clinical imaging/follow-up, and 1,988 participants with normal bone radiographs (1,152 males, 836 females; mean age 27.9±12.2 years). The dataset was split into a training set for model development, an internal independent and an external test set for model validation. The trained model located bone tumor lesions and then detected tumor patients. Receiver operating characteristic curves and Cohen's kappa coefficient were used for evaluating detection performance. We compared the model's detection performance with that of two junior radiologists in the internal test set using permutation tests. Results: The DL model correctly localized 94.5% and 92.9% bone tumors on radiographs in the internal and external test set, respectively. An accuracy of 0.964/0.920, and an area under the receiver operating characteristic curve (AUC) of 0.981/0.990 in DL detection of bone tumor patients were for the internal and external test set, respectively. Cohen's kappa coefficient of the model in the internal test set was significantly higher than that of the two junior radiologists with 4 and 3 years of experience in musculoskeletal radiology (Model vs. Reader A, 0.927 vs. 0.777, P<0.001; Model vs. Reader B, 0.927 vs. 0.841, P=0.033). Conclusions: The DL model achieved good performance in detecting primary bone tumors around the knee joint. This model had better performance than those of junior radiologists, indicating the potential for the detection of bone tumors on radiographs.
RESUMO
PURPOSE: To develop and validate a nomogram model that combines radiomics features, clinical factors, and coagulation function indexes (CFI) to predict intraoperative blood loss (IBL) during cesarean sections, and to explore its application in optimizing perioperative management and reducing maternal morbidity. METHODS: In this retrospective consecutive series study, a total of 346 patients who underwent magnetic resonance imaging (156 for training and 68 for internal test, center 1; 122 for external test, center 2) were included. IBL+ was defined as more than 1000 mL estimated blood loss during cesarean sections. The prediction models of IBL were developed based on machine-learning algorithms using CFI, radiomics features, and clinical factors. ROC analysis was performed to evaluate the performance for IBL diagnosis. RESULTS: The support vector machine model incorporating all three modalities achieved an AUC of 0.873 (95% CI 0.769-0.941) and a sensitivity of 1.000 (95% CI 0.846-1.000) in the internal test set, with an AUC of 0.806 (95% CI 0.725-0.872) and a sensitivity of 0.873 (95% CI 0.799-0.922) in the external test set. It was also scored significantly higher than the CFI model (P = 0.035) on the internal test set, and both the CFI (P = 0.002) and radiomics-CFI models (P = 0.007) on the external test set. Additionally, the nomogram constructed based on three modalities achieved an internal testing set AUC of 0.960 (95% CI 0.806-0.999) and an external testing set AUC of 0.869 (95% CI 0.684-0.967) in the pregnant population without a pernicious placenta previa. It is noteworthy that the AUC of the proposed model did not show a statistically significant improvement compared to the Clinical-CFI model in both internal (P = 0.115) and external test sets (P = 0.533). CONCLUSION: The proposed model demonstrated good performance in predicting intraoperative blood loss (IBL), exhibiting high sensitivity and robust generalizability, with potential applicability to other surgeries such as vaginal delivery and postpartum hysterectomy. However, the performance of the proposed model was not statistically significantly better than that of the Clinical-CFI model.
Assuntos
Perda Sanguínea Cirúrgica , Cesárea , Imageamento por Ressonância Magnética , Nomogramas , Humanos , Feminino , Gravidez , Estudos Retrospectivos , Adulto , Imageamento por Ressonância Magnética/métodos , Valor Preditivo dos Testes , Sensibilidade e EspecificidadeRESUMO
Purpose: Retinopathy of prematurity (ROP) is a retinal vascular proliferative disease common in low birth weight and premature infants and is one of the main causes of blindness in children.In the context of predictive, preventive and personalized medicine (PPPM/3PM), early screening, identification and treatment of ROP will directly contribute to improve patients' long-term visual prognosis and reduce the risk of blindness. Thus, our objective is to establish an artificial intelligence (AI) algorithm combined with clinical demographics to create a risk model for ROP including treatment-requiring retinopathy of prematurity (TR-ROP) infants. Methods: A total of 22,569 infants who underwent routine ROP screening in Shenzhen Eye Hospital from March 2003 to September 2023 were collected, including 3335 infants with ROP and 1234 infants with TR-ROP among ROP infants. Two machine learning methods of logistic regression and decision tree and a deep learning method of multi-layer perceptron were trained by using the relevant combination of risk factors such as birth weight (BW), gestational age (GA), gender, whether multiple births (MB) and mode of delivery (MD) to achieve the risk prediction of ROP and TR-ROP. We used five evaluation metrics to evaluate the performance of the risk prediction model. The area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUCPR) were the main measurement metrics. Results: In the risk prediction for ROP, the BW + GA demonstrated the optimal performance (mean ± SD, AUCPR: 0.4849 ± 0.0175, AUC: 0.8124 ± 0.0033). In the risk prediction of TR-ROP, reasonable performance can be achieved by using GA + BW + Gender + MD + MB (AUCPR: 0.2713 ± 0.0214, AUC: 0.8328 ± 0.0088). Conclusions: Combining risk factors with AI in screening programs for ROP could achieve risk prediction of ROP and TR-ROP, detect TR-ROP earlier and reduce the number of ROP examinations and unnecessary physiological stress in low-risk infants. Therefore, combining ROP-related biometric information with AI is a cost-effective strategy for predictive diagnostic, targeted prevention, and personalization of medical services in early screening and treatment of ROP.
RESUMO
BACKGROUND: Occult peritoneal metastases (OPM) in patients with pancreatic ductal adenocarcinoma (PDAC) are frequently overlooked during imaging. The authors aimed to develop and validate a computed tomography (CT)-based deep learning-based radiomics (DLR) model to identify OPM in PDAC before treatment. METHODS: This retrospective, bicentric study included 302 patients with PDAC (training: n =167, OPM-positive, n =22; internal test: n =72, OPM-positive, n =9: external test, n =63, OPM-positive, n =9) who had undergone baseline CT examinations between January 2012 and October 2022. Handcrafted radiomics (HCR) and DLR features of the tumor and HCR features of peritoneum were extracted from CT images. Mutual information and least absolute shrinkage and selection operator algorithms were used for feature selection. A combined model, which incorporated the selected clinical-radiological, HCR, and DLR features, was developed using a logistic regression classifier using data from the training cohort and validated in the test cohorts. RESULTS: Three clinical-radiological characteristics (carcinoembryonic antigen 19-9 and CT-based T and N stages), nine HCR features of the tumor, 14 DLR features of the tumor, and three HCR features of the peritoneum were retained after feature selection. The combined model yielded satisfactory predictive performance, with an area under the curve (AUC) of 0.853 (95% CI: 0.790-0.903), 0.845 (95% CI: 0.740-0.919), and 0.852 (95% CI: 0.740-0.929) in the training, internal test, and external test cohorts, respectively (all P <0.05). The combined model showed better discrimination than the clinical-radiological model in the training (AUC=0.853 vs. 0.612, P <0.001) and the total test (AUC=0.842 vs. 0.638, P <0.05) cohorts. The decision curves revealed that the combined model had greater clinical applicability than the clinical-radiological model. CONCLUSIONS: The model combining CT-based DLR and clinical-radiological features showed satisfactory performance for predicting OPM in patients with PDAC.
Assuntos
Carcinoma Ductal Pancreático , Aprendizado Profundo , Neoplasias Pancreáticas , Neoplasias Peritoneais , Tomografia Computadorizada por Raios X , Humanos , Neoplasias Peritoneais/diagnóstico por imagem , Neoplasias Peritoneais/secundário , Carcinoma Ductal Pancreático/diagnóstico por imagem , Carcinoma Ductal Pancreático/secundário , Carcinoma Ductal Pancreático/patologia , Masculino , Neoplasias Pancreáticas/diagnóstico por imagem , Neoplasias Pancreáticas/patologia , Feminino , Estudos Retrospectivos , Pessoa de Meia-Idade , Idoso , Adulto , RadiômicaRESUMO
BACKGROUND: Different placenta accreta spectrum (PAS) subtypes pose varying surgical risks to the parturient. Machine learning model has the potential to diagnose PAS disorder. PURPOSE: To develop a cascaded deep semantic-radiomic-clinical (DRC) model for diagnosing PAS and its subtypes based on T2-weighted MRI. STUDY TYPE: Retrospective. POPULATION: 361 pregnant women (mean age: 33.10 ± 4.37 years), suspected of PAS, divided into segment training cohort (N = 40), internal training cohort (N = 139), internal testing cohort (N = 60), and external testing cohort (N = 122). FIELD STRENGTH/SEQUENCE: Coronal T2-weighted sequence at 1.5 T and 3.0 T. ASSESSMENT: Clinical characteristics such as history of uterine surgery and the presence of placenta previa, complete placenta previa and dangerous placenta previa were extracted from clinical records. The DRC model (incorporating radiomics, deep semantic features, and clinical characteristics), a cumulative radiological score method performed by radiologists, and other models (including a radiomics and clinical, the clinical, radiomics and deep learning models) were developed for PAS disorder diagnosing (existence of PAS and its subtypes). STATISTICAL TESTS: AUC, ACC, Student's t-test, the Mann-Whitney U test, chi-squared test, dice coefficient, intraclass correlation coefficients, least absolute shrinkage and selection operator regression, receiver operating characteristic curve, calibration curve with the Hosmer-Lemeshow test, decision curve analysis, DeLong test, and McNemar test. P < 0.05 indicated a significant difference. RESULTS: In PAS diagnosis, the DRC-1 outperformed than other models (AUC = 0.850 and 0.841 in internal and external testing cohorts, respectively). In PAS subtype classification (abnormal adherent placenta and abnormal invasive placenta), DRC-2 model performed similarly with radiologists (P = 0.773 and 0.579 in the internal testing cohort and P = 0.429 and 0.874 in the external testing cohort, respectively). DATA CONCLUSION: The DRC model offers efficiency and high diagnostic sensitivity in diagnosis, aiding in surgical planning. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY: Stage 2.
RESUMO
Conventional radiomics analysis requires the manual segmentation of lesions, which is time-consuming and subjective. This study aimed to assess the feasibility of predicting muscle invasion in bladder cancer (BCa) with radiomics using a semi-automatic lesion segmentation method on T2-weighted images. Cases of non-muscle-invasive BCa (NMIBC) and muscle-invasive BCa (MIBC) were pathologically identified in a training cohort and in internal and external validation cohorts. For bladder tumor segmentation, a deep learning-based semi-automatic model was constructed, while manual segmentation was performed by a radiologist. Semi-automatic and manual segmentation results were respectively used in radiomics analyses to distinguish NMIBC from MIBC. An equivalence test was used to compare the models' performance. The mean Dice similarity coefficients of the semi-automatic segmentation method were 0.836 and 0.801 in the internal and external validation cohorts, respectively. The area under the receiver operating characteristic curve (AUC) were 1.00 (0.991) and 0.892 (0.894) for the semi-automated model (manual) on the internal and external validation cohort, respectively (both p < 0.05). The average total processing time for semi-automatic segmentation was significantly shorter than that for manual segmentation (35 s vs. 92 s, p < 0.001). The BCa radiomics model based on semi-automatic segmentation method had a similar diagnostic performance as that of manual segmentation, while being less time-consuming and requiring fewer manual interventions.
RESUMO
BACKGROUND AND OBJECTIVES: Radiomics and deep learning are two popular technologies used to develop computer-aided detection and diagnosis schemes for analysing medical images. This study aimed to compare the effectiveness of radiomics, single-task deep learning (DL) and multi-task DL methods in predicting muscle-invasive bladder cancer (MIBC) status based on T2-weighted imaging (T2WI). METHODS: A total of 121 tumours (93 for training, from Centre 1; 28 for testing, from Centre 2) were included. MIBC was confirmed with pathological examination. A radiomics model, a single-task model, and a multi-task model based on T2WI were constructed in the training cohort with five-fold cross-validation, and validation was conducted in the external test cohort. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic performance of each model. DeLong's test and a permutation test were used to compare the performance of the models. RESULTS: The area under the ROC curve (AUC) values of the radiomics, single-task and multi-task models in the training cohort were: 0.920, 0.933 and 0.932, respectively; and were 0.844, 0.884 and 0.932, respectively, in the test cohort. The multi-task model achieved better performance in the test cohort than did the other models. No statistically significant differences in AUC values and Kappa coefficients were observed between pairwise models, in either the training or test cohorts. According to the Grad-CAM feature visualization results, the multi-task model focused more on the diseased tissue area in some samples of the test cohort compared with the single-task model. CONCLUSIONS: The T2WI-based radiomics, single-task, and multi-task models all exhibited good diagnostic performance in preoperatively predicting MIBC, in which the multi-task model had the best diagnostic performance. Compared with the radiomics method, our multi-task DL method had the advantage of saving time and effort. Compared with the single-task DL method, our multi-task DL method had the advantage of being more lesion-focused and more reliable for clinical reference.
Assuntos
Aprendizado Profundo , Neoplasias da Bexiga Urinária , Humanos , Neoplasias da Bexiga Urinária/diagnóstico por imagem , Imageamento por Ressonância Magnética , Curva ROC , Músculos/diagnóstico por imagem , Estudos RetrospectivosRESUMO
OBJECTIVES: To compare the diagnostic performance of a novel deep learning (DL) method based on T2-weighted imaging with the vesical imaging-reporting and data system (VI-RADS) in predicting muscle invasion in bladder cancer (MIBC). METHODS: A total of 215 tumours (129 for training and 31 for internal validation, centre 1; 55 for external validation, centre 2) were included. MIBC was confirmed by pathological examination. VI-RADS scores were provided by two groups of radiologists (readers 1 and readers 2) independently. A deep convolutional neural network was constructed in the training set, and validation was conducted on the internal and external validation sets. ROC analysis was performed to evaluate the performance for MIBC diagnosis. RESULTS: The AUCs of the DL model, readers 1, and readers 2 were as follows: in the internal validation set, 0.963, 0.843, and 0.852, respectively; in the external validation set, 0.861, 0.808, and 0.876, respectively. The accuracy of the DL model in the tumours scored VI-RADS 2 or 3 was higher than that of radiologists in the external validation set: for readers 1, 0.886 vs. 0.600, p = 0.006; for readers 2, 0.879 vs. 0.636, p = 0.021. The average processing time (38 s and 43 s in two validation sets) of the DL method was much shorter than the readers, with a reduction of over 100 s in both validation sets. CONCLUSIONS: Compared to radiologists using VI-RADS, the DL method had a better diagnostic performance, shorter processing time, and robust generalisability, indicating good potential for diagnosing MIBC. KEY POINTS: ⢠The DL model shows robust performance for MIBC diagnosis in both internal and external validation. ⢠The diagnostic performance of the DL model in the tumours scored VI-RADS 2 or 3 is better than that obtained by radiologists using VI-RADS. ⢠The DL method shows potential in the preoperative assessment of MIBC.