BERT-based Transfer Learning in Sentence-level Anatomic Classification of Free-Text Radiology Reports.

Purpose: To assess whether transfer learning with a bidirectional encoder representations from transformers (BERT) model, pretrained on a clinical corpus, can perform sentence-level anatomic classification of free-text radiology reports, even for anatomic classes with few positive examples. Materials and Methods: This retrospective study included radiology reports of patients who underwent whole-body PET/CT imaging from December 2005 to December 2020. Each sentence in these reports (6272 sentences) was labeled by two annotators according to body part ("brain," "head & neck," "chest," "abdomen," "limbs," "spine," or "others"). The BERT-based transfer learning approach was compared with two baseline machine learning approaches: bidirectional long short-term memory (BiLSTM) and the count-based method. Area under the precision-recall curve (AUPRC) and area under the receiver operating characteristic curve (AUC) were computed for each approach, and AUCs were compared using the DeLong test. Results: The BERT-based approach achieved a macro-averaged AUPRC of 0.88 for classification, outperforming the baselines. AUC results for BERT were significantly higher than those of BiLSTM for all classes and those of the count-based method for the "brain," "chest," "abdomen," and "others" classes (P values < .025). AUPRC results for BERT were superior to those of baselines even for classes with few labeled training data (brain: BERT, 0.95, BiLSTM, 0.11, count based, 0.41; limbs: BERT, 0.74, BiLSTM, 0.28, count based, 0.46; spine: BERT, 0.82, BiLSTM, 0.53, count based, 0.69). Conclusion: The BERT-based transfer learning approach outperformed the BiLSTM and count-based approaches in sentence-level anatomic classification of free-text radiology reports, even for anatomic classes with few labeled training data.Keywords: Anatomy, Comparative Studies, Technology Assessment, Transfer Learning Supplemental material is available for this article. © RSNA, 2023.

Histogram-pattern analysis of the lung perfused blood volume for assessment of pulmonary thromboembolism.

PURPOSE: We aimed to evaluate the usefulness of histograms of lung perfused blood volume (HLPBV) based on the presence of pulmonary thromboembolism (PTE) and the pulmonary embolic burden. METHODS: A total of 168 patients (55 males; mean age, 62.9 years) underwent contrast-enhanced dual-energy computed tomography (DECT) between January 1 2012 and October 31 2014. Initial DECT images were three-dimensionally reconstructed, and the HLPBV patterns were divided into three types, including the symmetric type (131 patients, 78.0%), gradual type (25 patients, 14.9%), and asymmetric type (12 patients, 7.1%). RESULTS: Acute PTE was diagnosed in all 12 patients with asymmetric type (100%), 19 of the 25 patients with gradual type (76%) and 24 of the 131 patients with symmetric type (18.3%). HLPBV pattern exhibited correlations with the right/left ventricular diameter ratio (r=0.36, P = 0.007) and CT obstruction index (r=0.63, P < 0.001) in patients with PTEs. When the gradual and asymmetric types were regarded as positive for PTE, the specificity, positive predictive value, negative predictive value, and accuracy were 92.9%, 83.8%, 87.6%, and 81.0%, respectively. CONCLUSION: Histogram-pattern analysis using DECT might be a useful application to diagnose PTE.

Fractal analysis for the assessment of pulmonary blood flow on chest radiographs.

The authors investigated whether fractal analysis can be used with chest radiography for assessment of increased pulmonary blood flow. As a physical measure they used a fractal dimension (FD), which can be used for quantifying texture in selected regions of interest (ROIs). In general, the pulmonary capillary wedge pressure (PCWP) is normally less than 10 mm Hg, the redistribution of pulmonary blood flow reflects PCWP greater than or equal to 10 mm Hg, and the interstitial edema is observed in PCWP greater than or equal to 20 mm Hg. So, the study group consisted of 34 patients with mitral stenosis and 46 patients without mitral stenosis. These patients with and without mitral stenosis were divided into three groups based on their pulmonary capillary wedge pressure (PCWP): normal (n = 38; PCWP, <10 mmHg), mildly increased (n = 26; PCWP, 10 to <20 mmHg), and severely increased (n = 16; PCWP, >or=20 mmHg). ROIs were selected in the right upper and lower lung zones of digitized chest radiographs, and FDs were calculated in each ROI. As a measure of increased pulmonary blood flow, the upper-to-lower FD ratio (FDu/L) was calculated. Five chest radiologists estimated the degree of increased pulmonary blood flow for the same images of ROI sets presented on a cathode ray tube monitor in a randomized order. Between the normal PCWP group and the increased PCWP group, there were no significant differences in utility between the FDu/L and radiologists (P = 0.063). However, when the normal and the mildly increased PCWP groups were compared, the utility of the FDu/L was superior to that of radiologists (P = 0.042). These results indicate that fractal analysis is useful for quantitative estimation of increased pulmonary blood flow, especially in cases of mildly increased pulmonary blood flow.