Leveraging open dataset and transfer learning for accurate recognition of chronic pulmonary embolism from CT angiogram maximum intensity projection images

Figures

• 

  Fig. 1. A flow chart of the data selection process for the public dataset. Dashed arrows indicate exam exclusions without overlap with previous exclusions. PE, Pulmonary embolism; RV/LV, Right ventricular to left ventricular diameter ratio; QA, Quality assurance

• 

  Fig. 2. Data splits in phases 1–3. The block widths correspond approximately to the relative number of cases in the experiment A and the local sets. These test sets are not to be confused with the private and public test sets in the original RSPECT dataset (see text for details), for which we did not have ground truths available. Experiment A sets or splits were subsets of the corresponding sets or splits in experiment B

• 

  Fig. 3. The neural network architecture with DenseNet-121 as the frozen base model. Individual feature vectors were averaged and fed to the three-layer classification network. Lungs were processed separately, and the outputs were either taken individually or combined by taking the maximum of the positive class. Note that only the parameters trained by gradient descent were frozen in the base model, i.e., the batch statistics (the running mean and variance in the batch normalisation layers in the base model) were calculated during fine-tuning. PE, Pulmonary embolism

• 

  Fig. 4. ROC curves of the final individual (light grey) and ensemble models (black). LRmax is calculated by taking the maximum of exam’s left and right lung chronic pulmonary embolism predictions, whereas all is calculated by treating each lung separately. AUC, Area under the curve; LRmax, Maximum of the left and right lung prediction; ROC, Receiver operator characteristic

• 

  Fig. 5. Examples of the CNN prediction visualisations are shown in yellow for two true positive cases. Segmental and subsegmental arteries of the upper lobes are reduced in diameter and are smaller than the accompanying bronchi (arrows in (a)). Abrupt narrowing of the left superior lingular artery (arrow in (b)) and reduction of pulmonary artery diameter (arrowheads in (b)) are seen as signs of chronic pulmonary embolism

• 

  Fig. 6. An example of tortuous course vessels in the left lung of a chronic pulmonary embolism patient shown in coronal (a) and MIP images (b). The right lung is hidden in b