TL;DR
This paper introduces a novel deep-integrated model that combines 3D deep learning with morphological features to detect aortic dissection (AD) on non-contrast CT scans. This approach significantly outperforms a standalone deep learning model and shows comparable accuracy to radiologists, offering a promising tool for earlier and more accessible AD diagnosis.
What Problem Does This Paper Address?
Aortic dissection (AD) is a critical, life-threatening condition where early diagnosis and treatment are paramount. While CT angiography (CTA) is the gold standard for diagnosis, it has limitations: contrast agent risks (allergenicity, nephrotoxicity), availability issues in some regions, and the need for specialized staff. Non-contrast CT scans, however, are widely available, convenient, and pose lower risks. Despite these advantages, detecting AD on non-contrast CT is challenging for radiologists due to subtle imaging characteristics, leading to suboptimal sensitivity. This paper aims to bridge this gap by developing an automated system that leverages non-contrast CT scans for improved AD detection.
Methods
The researchers developed a two-stage AD detection model. First, a 3D deep learning (DL) model, based on the ResNet34 architecture, was trained to extract features from the aorta. The output of this DL model, a "DL score," was then used as an input feature.
The core innovation lies in the second stage: a "deep-integrated model" built using a Gaussian Naive Bayes (Gaussian NB) algorithm. This model combines the DL score with various morphological characteristics extracted from the aorta. These morphological features include:
- Aortic Maximum Diameters: Specifically, the maximum diameters of the ascending aorta (AC) and descending aorta (DC), categorized by thresholds (e.g., AC > 4 cm).
- General Morphological Features: A set of 14 features extracted using PyRadiomics, which capture aspects like surface area, volume, and shape descriptors.
The model was trained and validated using data from two medical centers, and its performance was evaluated against the DL model alone and the subjective assessments of experienced radiologists.
Datasets
The study utilized retrospective data from two medical centers in China:
- Internal Cohort (Peking Union Medical College Hospital - PUMCH): 341 patients (139 with AD, 202 without AD). This cohort was further split into:
- Training cohort: 238 patients
- Validation cohort: 35 patients
- Internal testing cohort: 68 patients
- External Testing Cohort (Shenzhen Second People's Hospital - SSPH): 111 patients (46 with AD, 65 without AD).
These datasets were used for training, validation, and independent testing of the developed model.
Key Insights and Ideas
- Synergy of DL and Morphology: The most significant insight is that combining deep learning's ability to capture subtle textural patterns with explicit morphological features dramatically improves AD detection performance compared to either approach alone. This suggests that a holistic view, integrating both "what it looks like" (DL) and "how it's shaped" (morphology), is crucial.
- Non-Contrast CT's Potential: The study strongly supports the viability of using non-contrast CT for AD detection, addressing a critical need where contrast-enhanced scans are not feasible or ideal.
- Radiologist Performance Baseline: The comparison with radiologists provides a crucial benchmark. While the model's accuracy was comparable, its significantly higher sensitivity in some cases suggests it could act as a valuable "second reader" or screening tool, potentially catching cases missed by human eyes.
- Feature Importance: The study highlights that while DL scores are important, specific morphological features (like aortic diameters) also play a significant role, particularly in differentiating AD from non-AD.
Key Results
- Superior Performance: The deep-integrated model achieved significantly higher AUCs than the 3D DL model alone:
- Internal testing cohort: AUC of 0.948 (vs. 0.803 for DL model)
- External testing cohort: AUC of 0.969 (vs. 0.814 for DL model)
- High Accuracy Metrics:
- Internal testing cohort: Accuracy 0.897, Sensitivity 0.862, Specificity 0.923
- External testing cohort: Accuracy 0.730, Sensitivity 0.978, Specificity 0.554
- Radiologist Comparison: The deep-integrated model showed comparable accuracy to radiologists (p > 0.05). However, its sensitivity was notably higher, especially in the external testing cohort (0.978 vs. radiologists' ~0.3-0.7), though its specificity was lower in that cohort.
Limitations and Future Work
- Specificity in External Cohort: A notable limitation is the lower specificity (0.554) in the external testing cohort compared to the internal cohort. The authors attribute this, in part, to differences in slice thickness and potentially other scanning parameters between the centers.
- Retrospective Nature: The study is retrospective, and detailed information on initial symptoms and the specific purpose of CT scans for all patients was incomplete.
- Generalizability: While validated across two centers, further multicenter prospective trials with diverse CT examination types are needed to fully establish the model's clinical potential and generalizability.
- Slice Thickness Impact: The study suggests that thicker slices in the external cohort might have impacted specificity, indicating a need for further investigation into how varying image acquisition parameters affect model performance.
- Clinical Integration: The authors suggest that the model could be more helpful for community radiologists lacking specialized cardiovascular imaging experience and propose future work on generating synthetic CTA from non-contrast CT, though they acknowledge the challenges of creating comprehensive datasets for this.
Summary updated