Machine learning for the prediction of pathologic pneumatosis intestinalis.

BACKGROUND: The radiographic finding of pneumatosis intestinalis can indicate a spectrum of underlying processes ranging from a benign finding to a life-threatening condition. Although radiographic pneumatosis intestinalis is relatively common, there is no validated clinical tool to guide surgical management. METHODS: Using a retrospective cohort of 300 pneumatosis intestinalis cases from a single institution, we developed 3 machine learning models for 2 clinical tasks: (1) the distinction of benign from pathologic pneumatosis intestinalis cases and (2) the determination of patients who would benefit from an operation. The 3 models are (1) an imaging model based on radiomic features extracted from computed tomography scans, (2) a clinical model based on clinical variables, and (3) a combination model using both the imaging and clinical variables. RESULTS: The combination model achieves an area under the curve of 0.91 (confidence interval: 0.87-0.94) for task I and an area under the curve of 0.84 (confidence interval: 0.79-0.88) for task II. The combination model significantly (P < .05) outperforms the imaging model and the clinical model for both tasks. The imaging model achieves an area under the curve of 0.72 (confidence interval: 0.57-0.87) for task I and 0.68 (confidence interval: 0.61-0.74) for task II. The clinical model achieves an area under the curve of 0.87 (confidence interval: 0.83-0.91) for task I and 0.76 (confidence interval: 0.70-0.81) for task II. CONCLUSION: This study suggests that combined radiographic and clinical features can identify pathologic pneumatosis intestinalis and aid in patient selection for surgery. This tool may better inform the surgical decision-making process for patients with pneumatosis intestinalis.

Development and validation of a five-factor score for prediction of pathologic pneumatosis.

BACKGROUND: The significance of pneumatosis intestinalis (PI) remains challenging. While certain clinical scenarios are predictive of transmural ischemia, risk models to assess the presence of pathologic PI are needed. The aim of this study was to determine what patient factors at the time of radiographic diagnosis of PI predict the risk for pathologic PI. METHODS: We conducted a retrospective cohort study examining patients with PI from 2010 to 2016 at a multicenter hospital network. Multivariate logistic regression was used to develop a predictive model for pathologic PI in a derivation cohort. Using regression-coefficient-based methods, the final multivariate model was converted into a five-factor-based score. Calibration and discrimination of the score were then assessed in a validation cohort. RESULTS: Of 305 patients analyzed, 102 (33.4%) had pathologic PI. We identified five factors associated with pathologic PI at the time of radiographic diagnosis: small bowel PI, age 70 years or older, heart rate 110 bpm or greater, lactate of 2 mmol/L or greater, and neutrophil-lymphocyte ratio 10 or greater. Using this model, patients in the validation cohort were assigned risk scores ranging from 0 to 11. Low-risk patients were categorized when scores are 0 to 4; intermediate, score of 5 to 6; high, score of 7 to 8; and very high risk, 9+. In the validation cohort, very high-risk patients (n = 17; 18.1%) had predicted rates of pathologic pneumatosis of 88.9% and an observed rate of 82.4%. In contrast, patients labeled as low risk (n = 37; 39.4%) had expected rates of pathologic pneumatosis of 1.3% and an observed rate of 0%. The model showed excellent discrimination (area under the curve, 0.90) and good calibration (Hosmer-Lemeshow goodness-of-fit, p = 0.37). CONCLUSION: Our score accurately stratifies patient risk of pathologic pneumatosis. This score has the potential to target high-risk individuals for expedient operation and spare low-risk individuals invasive interventions. LEVEL OF EVIDENCE: Prognostic Study, Level III.

Deep Learning Pre-training Strategy for Mammogram Image Classification: an Evaluation Study.

In this work, we assess how pre-training strategy affects deep learning performance for the task of distinguishing false-recall from malignancy and normal (benign) findings in digital mammography images. A cohort of 1303 breast cancer screening patients (4935 digital mammogram images in total) was retrospectively analyzed as the target dataset for this study. We assessed six different convolutional neural network model structures utilizing four different imaging datasets (total > 1.4 million images (including ImageNet); medical images different in terms of scale, modality, organ, and source) for pre-training on six classification tasks to assess how the performance of CNN models varies based on training strategy. Representative pre-training strategies included transfer learning with medical and non-medical datasets, layer freezing, varied network structure, and multi-view input for both binary and triple-class classification of mammogram images. The area under the receiver operating characteristic curve (AUC) was used as the model performance metric. The best performing model out of all experimental settings was an AlexNet model incrementally pre-trained on ImageNet and a large Breast Density dataset. The AUC for the six classification tasks using this model ranged from 0.68 to 0.77. In the case of distinguishing recalled-benign mammograms from others, four out of five pre-training strategies tested produced significant performance differences from the baseline model. This study suggests that pre-training strategy influences significant performance differences, especially in the case of distinguishing recalled- benign from malignant and benign screening patients.