.jpg)
In the field of diagnostic imaging in thoracic oncology, the comparison between foundational deep learning models and traditional radiomics has gained increasing interest. A recent study evaluates how different feature extractors, classifiers, and segmentation regimes influence tasks such as tumor volume classification, staging, survival prediction, or histological type. The results underscore that there is no single solution: the optimal design depends largely on the specific clinical task. For example, tumor segmentation is critical for predicting volume and stage, while the choice of classifier has a greater impact on survival and histology. This has direct implications for the development of tailored applications in hospital settings, where the integration of artificial intelligence for businesses requires balancing accuracy, robustness, and scalability.
Radiomics, despite its maturity, remains competitive in tasks such as tumor volume and staging, partly due to its relationship with labels derived from the segmentations themselves. In contrast, models like Curia and its variants achieve comparable scores in survival prediction, while DINOv3 lags slightly behind. A relevant finding is that patch- or slice-based aggregation methods have minimal impact, simplifying pipeline design. For research teams or companies seeking to implement robust solutions, we recommend exploring AI for businesses strategies that allow adapting these findings to small-sized cohorts, avoiding overfitting through two-stage designs.
From a technical perspective, the choice of classifier—from logistic regression to ensemble methods like CatBoost—can make a difference. The study's general recommendation is to use Curia with tumor segmentation and CatBoost as the classifier head, achieving the best average ranking across the three main clinical tasks. However, when precise tumor delineations are not available, a viable alternative is to combine Curia-2 with lung segmentation and logistic regression. This demonstrates the importance of having tailored applications that allow customizing each pipeline component according to project needs.
In this context, the role of a software development company like Q2BSTUDIO is key. Not only because we offer artificial intelligence services and AI agents adapted to sectors such as healthcare, but also because we integrate cybersecurity capabilities, AWS and Azure cloud services, and business intelligence services like Power BI. Our approach enables machine learning models to be deployed securely, scalably, and audited. For example, a hospital wishing to implement a CT-based lung cancer phenotyping system can benefit from custom software development that combines radiomics or foundational models with a robust cloud infrastructure and analytical dashboards.
The research also highlights that, although foundational models promise some generalization, robustness across cohorts remains a challenge. The study's main metric was worst-case performance across cohorts, reflecting the need to test systems on external populations. Here, Q2BSTUDIO's experience in AWS and Azure cloud services can help manage large volumes of imaging data and maintain regulatory compliance. Likewise, incorporating AI agents to automate clinical workflows—from segmentation to report generation—is a value-added line.
Ultimately, the comparison between foundational models and radiomics in lung CT reveals that there is no absolute winner, but rather the design depends on the task. What matters is having the technical flexibility and necessary knowledge to implement customized solutions. At Q2BSTUDIO, we offer that know-how, combining custom application development with artificial intelligence, cybersecurity, cloud, and business intelligence, so organizations can make the most of advances in medical imaging. If your company or institution is looking to take the leap toward precision medicine, exploring our capabilities in AI for businesses can be the first step.
