The steady increase in hospital admissions and requests for computed tomography (CT) examinations by emergency departments makes the need for rapid and dependable diagnosis, and communication of results, a pressing issue for radiologists. This is particularly true for suspected acute pulmonary embolism (PE), where early initiation of anticoagulant therapy is linked to improved outcomes [1]. PE is being recognized as a potentially life-threatening condition that needs accurate diagnosis. It ranks as the third most frequent acute cardiovascular event, with a high annual incidence [2], [3], [4]. It is a particularly severe event, followed by patient death in one out of ten cases within the first three months, and up to one out of three cases in cohorts followed over several years [5]. The diagnosis of PE is commonly based on clinical criteria, D-dimer dosage, and CT pulmonary angiography (CTPA) [4,6]. Furthermore, CTPA allows an early assessment of clinical severity by identifying the arterial obstruction index [7] and right ventricle (RV) dilatation [2]. Specifically, the Qanadli score (i.e, vascular obstruction index, which is an alternative method to assess PE severity), and the RV/left ventricle (LV) diameter ratio (i.e., the most validated CT biomarker to predict mortality from PE) are both crucial metrics for assessing the severity and impact of PE [8], [9], [10], [11], [12]. In particular, studies have shown the significance of the Qanadli score in predicting mortality, detecting acute PE, and providing valuable insights into patient prognosis [13,14]. However, manually computing the Qanadli score in clinical settings can be challenging and time-consuming due to its complex scoring process.

The emergence of artificial intelligence (AI), and specifically deep learning, has sparked interest in the development of automated diagnostic tools and reading-assistance tools in radiology [15], [16], [17]. Several works focused on the use of 2D, 2.5D and 3D U-Net [18] architectures on 2D CT, consecutive 2D CT slices and 3D inputs respectively [19], [20], [21], the use of hybrid network architectures such DenseNet [22] ResNet [23] with U-Nets leveraging transfer learning and attention mechanism [24] for 2D and 3D PE detection, segmentation, and classification. Moreover, sophisticated object detection-based methods were proposed, such as P-Mask R-CNN [25], Faster R-CNN [26] built upon the instance segmentation neural network R-CNN [27] for enhance small PE detection. However, all these studies only considered PE detection and discarded the integration of PE severity quantification scores such as the Qanadli score and the RV/LV diameter ratio within the PE detection for 3D examinations.

To address the complexity of manual calculations, automating the estimation of both the Qanadli score and the RV/LV diameter ratio using advanced machine and deep learning methods could propose promising solutions. Particularly, it could offer speed, reproducibility, and a reduction in interobserver variability, thus improving efficiency and precision in PE severity quantification.

In 2022, the French Society of Radiology (SFR) organized a data challenge where the primary objective was to assist the diagnosis of pulmonary embolism (PE) by identifying the presence of an embolism, and estimating the Qanadli score and RV/LV diameter ratio [28]. This challenge aimed to focus on the use of AI for PE diagnosis on real-world data with limited annotation (i.e., deprived of pixel-level annotations of blood clots). This could bridge the gap between scientific contributions and their use in clinical domain.

The purpose of this study was to develop an automated deep learning-based method to detect PE on 3D CTPA examinations and propose an automatic estimation technique of PE severity quantification scores using the Qanadli score and the RV/LV diameter ratio.