- A new artificial intelligence (AI) model uses a dual approach to simultaneously analyze different views of CT scans, resembling how doctors work, but without the need to switch between perspectives.
- Researchers trained the model on scans from healthy individuals and lung cancer patients to distinguish between normal tissue, benign changes, and malignant tumours.
- The approach may help to improve early detection of lung cancer, especially in cases where tumours are small and harder to identify.
- Although further validation is necessary before clinical use, the researchers suggest it could enhance diagnostic accuracy and efficiency.
Early diagnosis of lung cancer is crucial, as it significantly improves survival rates. Estimates suggest the 5-year survival can increase from roughly 10% in late stages to more than 90% in early stages.
The first step in diagnosing lung cancer is often through imaging tools, such as CT scans. However, diagnosing early stage lung cancer from CT scans can be challenging due to the small size of tumors, similarity to surrounding structures, and human error in interpretation.
Now, a study published in
Researchers at Kaunas University of Technology (KTU) designed an AI model that analyzes CT scans by simultaneously assessing both fine details and the broader anatomical context. This approach is intended to mirror how clinicians would interpret these medical images.
Traditionally, a radiologist would need to switch between views when reviewing CT images. But this process can be time consuming and may increase the risk of missing subtle details on the scan.
Thus, the AI system aims to overcome this limitation by integrating both perspectives into a single analytical process.
The research team suggest the AI model is capable of evaluating local features, such as small nodules, while also considering their position and significance within the whole lung.
In a press release, study author Inzamam Mashood Nasir, PhD, explained that “you can think of it as having a magnifying glass and a full view of the scan at the same time.”
To build the system, the team trained the AI model using CT scans from both healthy individuals and patients with lung cancer. This enabled the AI model to differentiate between normal tissue, benign changes, and malignant tumours.
The system achieved an accuracy of over 96%, outperforming existing approaches and maintaining stable performance across different tests.
This dual-scale learning approach could be particularly useful in identifying early stage lung cancer, when tumours are typically small and more difficult to detect.
Lung cancer remains a leading cause of cancer-related death worldwide, largely because it is often diagnosed at an advanced stage. Earlier detection is strongly associated with better outcomes, making improved screening tools a major focus of ongoing research.
“The potential impact is improved consistency and possibly earlier identification of suspicious findings, which may support earlier intervention,” Nasir told Medical News Today.
“However, the effect on detection rates and patient outcomes would still need prospective clinical validation,” he added.
AI-based systems are increasingly being explored to maintain accuracy and reduce variability in scan interpretation.
The KTU researchers suggest that their AI model could support clinicians by improving diagnostic accuracy, reducing the likelihood of missed lesions, and speeding up image analysis. This could also help reduce the number of false alarms, which can lead to unnecessary stress and procedures.
“In terms of clinical use, this would be best described as a decision-support or second-reader tool for radiologists, helping flag suspicious CT scans and supporting prioritization, rather than replacing clinical judgment,” said study author Eunchan Kim, PhD, to MNT.
However, the researchers note that the model was trained on a relatively limited dataset. They add that further testing in real-world settings is still necessary, particularly in larger, more diverse patient groups.
While still in the research phase and requiring clinical validation and real-world testing, the new model highlights the growing role of AI in medical imaging.
By closely replicating how doctors interpret scans, such systems may eventually become valuable tools for early lung cancer detection, potentially improving survival rates through earlier intervention.
“The main challenges before real-world use are generalizability, external validation, workflow integration, and broader clinical adoption,” study author Samia Nawaz Yousafzai, BSSE, told MNT.
“Our study used a relatively small dataset and did not include external validation on an independent cohort,” she nored.
The team also suggest that similar AI approaches could be applied to other medical imaging tasks that also require both detailed and contextual understanding, such as brain tumours, breast cancer, and eye diseases.
“The natural next steps would be testing on larger multi-center datasets and collaborating with hospitals and radiology departments for prospective or real-time validation,” concluded Nasir.
