TL;DR
This paper introduces a deep learning-based computer-aided diagnostic (DL-CAD) system designed to improve the accuracy and efficiency of diagnosing rib fractures from CT scans in patients with blunt chest trauma. The study demonstrates that this DL-CAD system significantly enhances diagnostic performance, reduces reading time, and boosts radiologist confidence, particularly for less experienced practitioners.
What Problem Does This Paper Address?
Acute rib fractures (ARFs) are a common consequence of blunt chest trauma, but their diagnosis can be challenging. While CT scans are the most sensitive imaging modality, identifying subtle or non-displaced fractures can be time-consuming and error-prone, even for experienced radiologists. This diagnostic difficulty is exacerbated by the increasing volume of medical imaging data and a slower growth rate in the number of radiologists. Missed or delayed diagnoses of ARFs can have significant clinical implications, affecting patient treatment, forensic disability classification, and potentially leading to unnecessary medical disputes. Existing deep learning applications in medical imaging have focused on areas like pulmonary nodules or tumors, but their specific value in improving rib fracture diagnosis, especially for radiologists with varying experience levels, needed further validation.
Methods
The study retrospectively analyzed CT images from 214 patients with acute blunt chest trauma. Two groups of radiologists—two interns (1 year of experience) and two attending radiologists (over 7 years of experience)—independently interpreted the CT scans in a blinded and randomized manner. This was followed by a second reading session, where the same radiologists re-interpreted the scans with the assistance of a DL-CAD system. The DL-CAD system used an artificial convolutional neural network to automatically detect rib fractures and provided an information list with bounding boxes around suspected fractures.
Key performance metrics evaluated included:
- Diagnostic Sensitivity: The ability to correctly identify patients with rib fractures.
- Specificity: The ability to correctly identify patients without rib fractures.
- Positive Predictive Value (PPV): The proportion of positive diagnoses that are actually correct.
- Diagnostic Confidence: Assessed using a 5-point Likert scale for each rib.
- Reading Time: The time taken to interpret each patient's scans.
A consensus diagnosis by two senior thoracic radiologists served as the ground truth. Statistical analyses, including ROC analysis, paired t-tests, and chi-square tests, were used to compare performance metrics with and without DL-CAD assistance.
Datasets
The study utilized a dataset of 214 retrospective chest CT examinations from patients diagnosed with acute blunt chest trauma. These scans were collected from July 2018 to February 2019. The ground truth for rib fractures was established by two senior radiologists. The dataset contained a total of 680 confirmed rib fracture lesions, comprising 529 complete rib fractures (CRFs) and 151 occult rib fractures (ORFs).
Key Insights and Ideas
- DL-CAD as an Equalizer: The DL-CAD system significantly boosted the diagnostic performance of interns, bringing their sensitivity and PPV close to that of attending radiologists. This suggests DL-CAD can help bridge the experience gap between junior and senior radiologists.
- Improved Detection of Occult Fractures: A notable improvement was observed in the detection of occult rib fractures (ORFs), which are notoriously difficult to identify. DL-CAD assistance led to a substantial increase in ORF sensitivity for both interns and attendings.
- Confidence Boost: Radiologists, regardless of experience, reported higher diagnostic confidence when using the DL-CAD system. This is crucial for making timely and definitive treatment decisions.
- Efficiency Gains: The DL-CAD system dramatically reduced the time required for radiologists to interpret scans, freeing up valuable clinical time.
Key Results
- Interns' Performance: With DL-CAD assistance, interns' sensitivity increased from 68.82% to 91.76% (P<0.05) and PPV from 84.50% to 93.17% (P<0.05). ORF sensitivity improved from 47.68% to 80.13%.
- Attendings' Performance: DL-CAD assistance improved attendings' sensitivity from 86.47% to 94.56% (P<0.05). ORF sensitivity increased from 74.17% to 86.09%.
- Diagnostic Confidence: Average diagnostic confidence scores increased significantly for both interns (from 3.69±1.12 to 4.32±0.87) and attendings (from 4.11±1.03 to 4.54±0.72). The proportion of "definitely present" ratings rose from 26.47% to 53.38% for interns and from 44.26% to 63.53% for attendings.
- Reading Time Reduction: Interns' reading time decreased by 53.4% (from 99.48s to 46.40s), and attendings' reading time decreased by 34.0% (from 65.96s to 43.54s) with DL-CAD assistance.
- ROC Analysis: The Area Under the Curve (AUC) for ROC analysis improved for both interns (from 0.925 to 0.977) and attendings (from 0.955 to 0.987) when using DL-CAD.
Limitations and Future Work
The authors acknowledge several limitations:
- Retrospective Design: The retrospective nature of the study might lead to an underestimation of missed diagnoses due to a lack of follow-up data for some patients.
- Small Sample Size and Single Center: The study used a relatively small sample size from a single clinical center, and all CT images were acquired from a GE scanner. Future work should validate the DL-CAD system on larger, multi-center datasets and across different scanner vendors.
- Gold Standard: The gold standard for fracture diagnosis was based on expert radiologist review, as few patients underwent surgical treatment. This could introduce some subjectivity.
Future work could focus on prospective studies to further validate these findings and explore the integration of DL-CAD systems into routine clinical workflows.
Summary updated