Second Opinion, Same Image: How 'Similar Subtraction' Could Change Lung Nodule Detection
"When a past X-ray isn't available, a new technique using images from other patients shows promise in spotting lung nodules."
Catching respiratory diseases early is crucial, and digital chest X-rays (CXRs) are a key tool. They're quick, cost-effective, and expose patients to less radiation than CT scans. But spotting subtle abnormalities on a CXR can be challenging. That's where techniques like temporal subtraction (TS) come in, which compares current and past images to highlight changes.
However, TS requires previous CXRs, which aren't always available. To address this, researchers have developed 'similar subtraction' (SS). Instead of relying on a patient's past images, SS uses CXRs from other patients with similar lung shapes. The question is: how well does it work compared to traditional methods?
A new study published in Radiological Physics and Technology dives into this question. Researchers compared SS images with TS images to see how well each technique could depict simulated lung nodules. This article breaks down their findings and what it could mean for the future of lung disease detection.
Current Imaging and Its Limits
Chest radiography is described as a first-line test for detecting lung nodules, but interpretation can be difficult because nodules may be small and low contrast against the complex thoracic region. CT is also used for nodule detection and follow-up; one clinical study protocol plans to enroll 150 patients referred for non-enhanced chest CT and add an unenhanced ultra-low-dose CT acquisition, with an average dose about 10 times lower than conventional low-dose CT. Even with imaging, many nodules remain indeterminate and may require follow-up over time to assess stability or invasive assessment for a definitive diagnosis.
Similar Subtraction vs. Temporal Subtraction: A Head-to-Head Comparison
The study involved 100 patients. The most recent CXR for each patient was selected and a simulated nodule was added to mimic the presence of a potential lung abnormality. The goal was to see how well the SS technique could reveal this nodule, especially when compared to the TS method.
- Temporal Subtraction (TS): This traditional method subtracted a previous CXR from the current one. Two intervals were used: 2-year and 7-year gaps between the images.
- Similar Subtraction (SS): This newer technique subtracted a 'similar' CXR (from a different patient) from the current one. The 'Top 1' most similar image and the 'Top 10' most similar images were tested.
- Image Processing: A non-linear image-warping technique was applied in both TS and SS to enhance alignment and reduce artifacts.
Detection Does Not Resolve Every Nodule
A study of AI-based detection software in routine chest radiographs evaluated how often clinically significant lung nodules were found incidentally and whether accompanying findings could help with differential diagnosis. Its report says AI-based software can help identify clinically significant incidental nodules, while referring to co-existing lung lesions may help classify them. These findings describe potential assistance with detection and classification, rather than establishing that software alone resolves every nodule. A separate survey reviews strengths and limitations across imaging-modality approaches to lung cancer detection.
Subtraction and Imaging Comparisons
In a study of standardized radiographs, subjects detected changes in alveolar bone defects with bone loss in the range of 1–5 percent using subtraction radiography, as measured by 125I absorptiometry. A separate periodontal-therapy study found that only one half of the radiographs were appropriate for subtraction analysis with the technology then available. For chest imaging, a meta-analysis was designed to compare digital tomosynthesis with chest radiography for pulmonary-nodule detection. These sources cover different settings and do not establish that subtraction radiography improves lung-nodule detection.
The Verdict: Can 'Similar Subtraction' Replace Past X-Rays?
The study revealed promising results for similar subtraction. In a portion of cases, the Top 1 SS images showed a higher CNR than the TS images. Specifically, 28% of Top 1 SS images outperformed TS images with a 2-year interval, and 33% outperformed TS images with a 7-year interval. Furthermore, when considering the Top 10 SS images, these percentages jumped to 56% and 72%, respectively.
Datasets and AI Development
The LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative) is a large, publicly available collection of thoracic CT scans annotated with lung nodules, intended to support the development and evaluation of computer-aided diagnostic methods. For chest X-rays, manual labeling is described as time-consuming and prone to errors and omissions, while nodule-size diversity and low contrast remain technical challenges. In a study of 10,476 people who had chest X-rays at a health screening center, AI assistance was associated with a higher detection rate for actionable lung nodules. A radiology review also reports that FDA-approved AI tools are already used clinically for tasks including nodule detection and expects continued increases in the breadth and accuracy of detection algorithms.
Data Gaps and Detection Challenges
Deep-learning methods have demonstrated high performance in studies of lung-nodule detection on chest X-rays, but the NODE21 source says the lack of gold-standard public datasets slows research progress. Chest radiography is described as economical and easy to use, yet radiologists may miss lung cancers visible in retrospect, and small lung cancers are especially difficult to detect. These sources point to both dataset availability and image interpretation as challenges for improving detection.
Training for Difficult Nodules
A 2024 study on chest-radiograph nodule detection notes that small, hard-to-detect nodules in large radiographic datasets are rarely validated by computed tomography. The study identifies these difficult nodules as important for training detection methods. It proposes artificial nodule synthesis algorithms that create artificially embedded nodules as a way to address the lack of such training examples. This source describes a training approach, not evidence that synthetic nodules have already improved patient outcomes.
This suggests that in a significant number of cases, SS can provide comparable or even better nodule depiction than TS, especially when long intervals exist between a patient’s current and previous CXRs.
While the study indicates the potential of SS, the researchers emphasize that it shouldn't replace traditional diagnostic methods. Instead, SS could serve as a valuable supplementary tool, increasing confidence in diagnoses and guiding decisions about further examinations, such as CT scans. Further research is needed to explore different types of nodules, and varied image qualities to fine-tune the SS technique and fully understand its capabilities.