As far as we know, this is the first study to evaluate the feasibility of three-dimensional CNN in the detection of lung hypoperfusion from CTPA images in CPE. Our CNN algorithm showed an encouraging performance in detecting these lesions on a separate testing set. Since there are no previous studies to compare our results to, we compared the CNN algorithm to a method based on a simple HU cutoff threshold. Using even the HU threshold segmentation could help visual evaluation, but according to our results, the CNN algorithm could offer a significant improvement with a 0.87 ROC-AUC and 0.46 MCC compared to the 0.79 and 0.35 for the naïve HU-threshold method, respectively.
The average CNN prediction probability indicates the presence of CPE, and all but one of the CPE and control cases in the test set were correctly categorised using a probability value 0.55. Therefore, we propose that in addition to the lesion segmentation, the disease overall presence could be inferred using a CNN-based approach.
The CNN algorithm recognised most (97%) of the manually segmented lesions. Majority (82%) of the false positive CNN labels were small and related to beam hardening artefacts from dense organs, which seemed to be the most challenging for the CNN to differentiate from true positive lesions. However, one patient had extensive false positive labels covering most of the lung volume. This patient had undergone left upper lobe pneumonectomy and mean HU of the lung parenchyma was one of the lowest in the study group. After lobectomy, the remaining lung has a relative increase of air and a decrease of the lung parenchyma in the thorax seen as hyperlucency on a CT image . This is likely the reason for the hyperlucency in this patient’s lung and the CNN false positive labelling.
There was both overestimation and underestimation in the sizes of the hypoperfused lesions. The overestimation was partly related to the beam hardening artefacts mentioned earlier, but in most cases, the overestimation and underestimation remained unspecific. Since accurate delineation of the hypoattenuating regions in the CTPA was challenging, we had to compromise the manual segmentation in cases where the borders between affected and unaffected areas were gradient. This might partially explain the overestimation and underestimation of the lesion sizes in the labels output by the CNN algorithm.
CNNs have been previously applied in pulmonary embolism detection based on labelling the occluding clots seen in CTPA as filling defects of intravascular contrast material in pulmonary arteries [39,40,41,42]. These studies have focused mainly on acute pulmonary embolism. We took a different approach, and instead of the actual vascular defects, we focused on the hypoattenuation in the lung parenchyma relating to the hypoperfusion caused by chronic pulmonary embolism. Öman et al. (2019) have successfully implemented this type of approach to detect acute stroke lesions from CT angiography images with a CNN based algorithm . However, since the lungs are composed largely of air, the differences in parenchymal enhancement are subtler between the hypo- and hyperperfused regions at CTPA than at CT angiography of the brain. Another significant difference in the setting is that a single thrombus usually causes the ischemic stroke, whereas chronic pulmonary embolism affects many branches of the pulmonary vasculature by various degrees of obstruction. Unlike the brain, the attenuation of the lung is sensitive to regional abnormalities of perfusion or aeration, which can be seen as heterogeneous lung attenuation on CT .
In patients with CPE, the areas of low attenuation on CTPA represent areas of hypoperfusion , which is due to both the obstructing vascular impairment caused by chronic thrombus and the vasculopathy in the regions clear of the thrombus . The low attenuation may also be due to air trapping in CTEPH patients relating to abnormalities in the affected regions' small airways . Our study had only CTPA studies without expiration images, so we could not definitely differentiate the hypoperfusion and air trapping. However, regardless of the cause, hypoattenuation is associated with the regions affected by CTE, and, i.e., Bartalena et al.  showed that the hypoattenuating regions seen in CTPA correlated well with hypoperfusion seen in perfusion scintigraphy of patients with pulmonary hypertension including CTEPH.
Absence of mosaic perfusion is not exclusive for CPE as in some studies up to 26–45% of patients with CTEPH did not present mosaic perfusion at CTPA [12, 48]. The mosaic perfusion is also not specific for CPE, and various diseases mimic this pattern . Hence, this type of CNN algorithm would not be a comprehensive tool for the diagnosis or exclusion of CPE, but it may help assess the extent of the disease and treatment planning. With a processing time well below 1 min, the CNN model could be implemented as a low-latency clinical application in the picture archiving and communication system by overlaying the prediction on top of the images or by reporting the volume(s) of the hypoperfused regions. This type of application might also assist the radiologist in diagnosing CPE, which is often misdiagnosed and difficult to detect, especially if combined with an algorithm assessing the vascular defects. A more extensive database is needed to conclude the feasibility of CNN algorithms in a clinical setting and with automatic methods of detecting the vascular defects.
A limitation of our study was that the labels were manually segmented by a single radiologist and contained an unknown interobserver variability and individual bias level. Contributing to this label noise was the lack of expiration CT scans, and definite differentiation between hypoperfusion and air trapping was not possible. Additionally, borders between the regions with differing perfusion were gradient in many cases, and exact demarcation was not always univocal during the pre-processing, which might explain some of the overestimation and underestimation of the lesions by the CNN algorithm.
The patients in the study had different parenchymal changes relating to other diseases (e.g., scars, atelectasis, pleural fluid, emphysema, inflammatory lesions), and all the patients who had these types of findings extended to over two thirds of the lung volume were excluded from the study. Although this exclusion criterion presents an unknown amount of bias in the sample profile, we found it necessary. Otherwise, the manual segmentation could not have been done objectively, leaving the segmentations too imprecise. To avoid inaccurate labelling, the smallest regions of hypoperfusion and regions with undefined borders relating to noise and artefacts were left manually unlabelled. The segmentation was done correspondingly for the whole group A (divided for the training, validation and the test sets). This had an unknown effect on the results as the CNN labelled some of these small regions correctly as hypoperfusion. Still, they presented as false positives since the manual segmentation was considered the ground truth. New practical methods for improved manual segmentation and labelling the parenchymal lesions in the lungs need to be explored. One such tool is the morphological contour interpolation introduced by Zukić et al. , which implements a morphology-based interslice interpolation proposed by Albu et al. . In our study, this tool turned out invaluable for the manual segmentation since the lesions were large in number and volume with irregular shapes and borders, and not always well demarcated in every CT slice.
Finally, we acknowledge as a study limitation the small sample size, for which only a modest anatomical and pathological variation were available during training. Also, the small number of validation cases may have led to a suboptimal stopping point and model selection. Nevertheless, to address the resulting problem of unknown generalizability, the data portion dedicated to testing was kept relatively high in the study design. Likely, a big data approach (e.g., training with additional, more readily available, non-CPE cases) would be beneficial, especially in decreasing the extent and frequency of false-positive regions.
In conclusion, this study demonstrated the feasibility of a deep learning algorithm for the detection of hypoperfusion in CPE from CTPA with a good performance. These encouraging findings suggest that CNNs could be used as an automated method for assisting the clinician and radiologist in diagnosis and treatment planning for patients with CPE.