Enhanced classification performance using deep learning based segmentation for pulmonary embolism detection in CT angiography

Figures

• 

  Fig. 1. Internal and external datasets for training and evaluation of a segmentation-based classification model for pulmonary embolism detection. Positive examinations refer to the patient having pulmonary embolism (PE) and negative examinations are patients without PE. True filling defect refers to tumor invasion, stump thrombus, catheter, embolized wire, or other obvious non-PE condition as defined in the RSPECT dataset.

• 

  Fig. 2. Training and evaluation of a segmentation-based classification model for pulmonary embolism detection. 700 CTPA examinations were collected and annotated by either of two radiologists. Of these, all 149 PE positive examinations were used for training and the PE negative were kept for later evaluation (a). The 3D U-Net deep learning model, which is generated by the nnU-Net framework, was trained with the 149 PE-positive CTPAs using 5-fold cross-validation. The convolution layer used a 3 × 3 × 3 filter size by default, followed by an instance normalization (IN) layer and a leaky rectified Linear Unit (lRELU) layer (b). The softmax probabilities were obtained from the model inference for fine-tuning classification into PE or non-PE voxel classes and for calculating the predicted PE volume. By thresholding the predicted volumes and applying a set of logical rules, accurate patient-level classification for PE was achieved (c). The final model was evaluated on 804 external CTPA examinations from two publicly available datasets (d).

• 

  Fig. 3. Classification performance of the trained nnU-Net model. Area Under the Curves (AUC) without (a) and with (b) post-processing. Black, internal dataset (n = 679, 128 PE and 551 non-PE); Blue, the FUMPE datasets (n = 34, 32 PE and 2 non-PE); Red, the RSNA PE dataset (n = 770, 385 PE and 385 non-PE). TPR, true positive rate; FPR, false positive rate. AUC values are in percentages.

• 

  Fig. 4. Representative segmentation results of the trained model. Axial, coronal, and sagittal planes from the same CTPA examinations from the external FUMPE dataset with the same window setting (width = 800 HU, level = 100 HU) are shown. Red, pulmonary embolism annotation; blue, model segmentation; purple, overlay of annotation and model segmentation.

• 

  Fig. 5. An illustration of a segmented reference pulmonary artery with reference threshold volumes. Patient orientation of 3D volumes with reference threshold volumes (20, 50, 200, 1000, and 10000 mm3) (a). Manually segmented reference pulmonary artery (volume of 113 cm³) from a male patient without PE (b). All volumetric images are isotropic (1 mm × 1 mm × 1 mm).