AI Shows Strong Correlation with Pathology in Diagnosing Multiple Primary Lung Cancers
In a compelling demonstration of artificial intelligence’s (AI) growing role in clinical oncology, a new study published in China Medical Devices reveals that AI-assisted CT analysis can reliably mirror pathological findings in patients with multiple primary lung cancer (MPLC). The research, conducted by a team from Peking University Third Hospital Haidian Campus and Infervision Technology Co., Ltd., underscores the potential of AI not only as a screening tool but as a clinically meaningful adjunct in the nuanced diagnosis of complex lung malignancies.
Multiple primary lung cancer—defined as the presence of two or more independent primary lung tumors in the same patient—poses a significant diagnostic challenge. Unlike metastatic disease, where secondary tumors originate from a single primary source, MPLC involves distinct cancerous lesions, each with its own biological behavior, prognosis, and therapeutic implications. Misdiagnosis can lead to inappropriate treatment strategies, underscoring the critical need for precise, early differentiation.
Historically, distinguishing MPLC from intrapulmonary metastases or recurrent disease has relied heavily on histopathological examination, often supplemented by molecular profiling and detailed imaging analysis. However, with the increasing adoption of low-dose CT screening programs worldwide, radiologists are encountering more incidental pulmonary nodules, many of which are small, ground-glass opacities (GGOs) that defy straightforward classification. This diagnostic ambiguity is especially pronounced in early-stage disease, where morphological features may be subtle.
Enter AI. Over the past decade, deep learning algorithms have demonstrated remarkable accuracy in detecting and characterizing pulmonary nodules on CT scans. But while much of the literature has focused on sensitivity and specificity in nodule detection, fewer studies have rigorously evaluated how well AI-derived metrics align with gold-standard pathological staging—particularly in multifocal lung cancer scenarios.
This new study bridges that gap. Led by Dasheng Li from the Department of Radiology at Peking University Third Hospital Haidian Campus, in collaboration with Dawei Wang from Infervision’s Institute of Advanced Research, the team retrospectively analyzed 57 confirmed cancerous lesions from 26 patients diagnosed with MPLC between 2017 and 2019. All patients underwent surgical resection—primarily via video-assisted thoracoscopic surgery (VATS)—and each lesion was independently validated by pathology as a true primary tumor, excluding metastatic or satellite lesions.
The researchers employed Infervision’s InferRead CT Lung, an AI-powered diagnostic support system, to automatically quantify key nodule characteristics: volume, longest diameter, shortest diameter, and density (solid vs. ground-glass). These AI-generated measurements were then correlated with pathological TNM staging according to the 8th edition guidelines—specifically stages 0, T1a1, T1a2, T1a3, T1b, T3a, and T3b.
The results were striking. Statistical analysis revealed a highly significant correlation (p < 0.001) between pathological stage and AI-measured nodule volume, longest diameter, and shortest diameter. As tumors progressed from ultra-early (stage 0/T1a1) to more advanced T categories, the AI system consistently recorded larger dimensions and greater volumes. This linear relationship held even after adjusting for tumor area, suggesting that AI metrics capture intrinsic biological progression beyond simple two-dimensional size.
Moreover, the study found that the AI system’s prediction of nodule type—solid versus ground-glass—was significantly associated with pathological staging (p < 0.05). Ground-glass nodules predominated in the early-stage group (80.6% in stage 0/T1a1), while solid nodules were markedly more common in the later-stage group (52.4% in T1a2–T3b). This aligns with established radiological-pathological correlations: ground-glass opacities often correspond to pre-invasive or minimally invasive adenocarcinomas, whereas solid components typically indicate invasive disease.
Perhaps most clinically relevant was the finding that nodule area ≥1.0 cm² was strongly predictive of higher pathological stage (p < 0.001). This quantitative threshold, automatically derived by AI, could serve as a practical red flag for radiologists evaluating multifocal lung lesions.
The implications extend beyond diagnostic accuracy. In an era of personalized oncology, timely identification of MPLC—rather than misclassifying it as metastatic disease—can dramatically alter management. Patients with true MPLC are often candidates for curative-intent resection of all lesions, whereas those with metastases may be directed toward systemic therapy. Thus, an AI tool that enhances the radiologist’s ability to suspect MPLC early could directly impact survival outcomes.
Critically, the study design reflects real-world clinical conditions. CT scans were acquired on standard commercial scanners (GE 64-slice VCT and Philips 128-slice iCT) using routine protocols, including both standard-dose and low-dose settings. This enhances the generalizability of the findings across diverse healthcare settings—not just academic centers with specialized imaging protocols.
The statistical rigor further bolsters credibility. Using SPSS 20.0, the team performed chi-square tests for categorical associations, Pearson correlation for continuous variables, and multivariate linear regression to model dependencies. In regression models, pathological stage independently predicted increases in longest diameter (β = 0.12, p = 0.01), while tumor area independently predicted both volume and longest diameter (p < 0.02 and p < 0.001, respectively). Adjusted R² values of 0.47 and 0.61 indicate moderate-to-strong explanatory power—remarkable for a purely imaging-based model in a heterogeneous disease like MPLC.
Notably, the study also addressed potential confounders. Tumor laterality (left vs. right lung) and timing of lesion discovery (synchronous vs. metachronous) showed no significant association with pathological stage, confirming that the observed AI-pathology correlations were not artifacts of anatomical or temporal bias.
While the sample size (26 patients, 57 lesions) is modest, it is comparable to other surgical-pathological validation studies in rare oncological subtypes. The authors acknowledge this limitation but argue that the granularity of pathological staging—spanning seven distinct T categories—adds analytical depth uncommon in similar investigations.
Looking ahead, the integration of AI into MPLC workflows could evolve in several directions. First, AI systems might incorporate not just size and density but also texture analysis, spiculation, pleural retraction, and vascular convergence—features known to correlate with malignancy. Second, future iterations could fuse imaging data with genomic or proteomic markers to predict not only stage but also driver mutations (e.g., EGFR, ALK), as hinted by prior Infervision-led research.
Third, and perhaps most transformative, AI could facilitate longitudinal monitoring. By precisely tracking subtle changes in nodule volume over time—even sub-millimeter shifts invisible to the human eye—AI might detect early transformation from indolent to aggressive phenotypes, enabling timely intervention.
From a health systems perspective, such technology could alleviate radiologist burnout. The exponential growth in CT utilization has created a tsunami of images requiring interpretation. AI tools that triage cases, flag high-risk nodules, and provide quantitative baselines allow radiologists to focus cognitive resources on complex decision-making rather than repetitive measurement tasks.
Regulatory bodies are taking note. The U.S. FDA and China’s NMPA have already cleared several AI-based pulmonary nodule detection systems. However, most are labeled for “detection assistance” only. This study moves the needle toward “diagnostic inference”—a higher regulatory bar that requires robust clinical validation against pathological outcomes, precisely what the authors have delivered.
Ethically, the deployment of such AI must be transparent. Radiologists should understand the algorithm’s training data, performance metrics, and failure modes. Importantly, AI should augment—not replace—clinical judgment. As the authors emphasize, their system is a “decision support” tool, intended to highlight suspicious features for expert review.
The broader oncology community stands to benefit. If AI can reliably stratify MPLC lesions by pathological aggressiveness, it could inform surveillance intervals, surgical planning (e.g., segmentectomy vs. lobectomy), and even adjuvant therapy decisions. For instance, a patient with multiple small ground-glass nodules might undergo staged resections or active surveillance, whereas one with a dominant solid nodule and smaller GGOs might receive more aggressive initial treatment.
This work also contributes to the global effort to standardize MPLC diagnosis. Current guidelines rely on a mix of clinical, radiological, and pathological criteria, but inter-observer variability remains high. An objective, quantitative AI framework could introduce much-needed consistency—especially in community hospitals lacking subspecialty thoracic radiologists.
In conclusion, the study by Li, Wang, Huang, Liu, and Huo represents a significant step toward evidence-based integration of AI into thoracic oncology. By demonstrating strong concordance between AI-derived CT metrics and surgical pathology across a spectrum of MPLC stages, the team provides compelling data that AI is not merely a novelty but a clinically valuable partner in the fight against lung cancer.
As low-dose CT screening expands globally—driven by landmark trials like NLST and NELSON—the incidence of MPLC will likely rise. Equipping clinicians with intelligent tools to navigate this complexity isn’t just advantageous; it’s essential. This research lights the path forward.
Authors: Dasheng Li¹ᵃ, Dawei Wang², Yuqing Huang¹ᵇ, Xiaoxu Liu¹ᵃ, Zhiyi Huo¹ᵃ
Affiliations:
¹ᵃ Department of Radiology, ¹ᵇ Department of Thoracic Surgery, Beijing Haidian Section of Peking University Third Hospital (Beijing Haidian Hospital), Beijing 100080, China
² Institute of Advanced Research, Infervision Technology Co., Ltd., Beijing 100025, China
Journal: China Medical Devices, 2021, Vol. 36, No. 02, pp. 77–80
DOI: 10.3969/j.issn.1674-1633.2021.02.019