RadiomicNet: Hybrid Radiomics-Deep Learning for Interpretable Segmentation

In the field of medical image segmentation, artificial intelligence has shown enormous potential, but faces challenges such as lack of interpretability and high computational consumption. RadiomicNet emerges as an innovative alternative: a hybrid architecture that combines deep neural networks with traditional radiomics features, achieving a balance between precision and transparency. This model, which integrates an attention module based on co-occurrence matrices and local binary patterns, not only improves performance on datasets like BUSI and Kvasir-SEG, but also drastically reduces the required parameters. The trend toward lighter, more explainable models is key in the development of AI for companies that require robust and auditable solutions. At Q2BSTUDIO, we apply these principles in image segmentation projects, adapting hybrid architectures for healthcare and industrial sectors. Our approach combines artificial intelligence with deep domain knowledge, offering custom software that integrates ante-hoc interpretability modules, essential for meeting clinical regulations. Additionally, the implementation of these systems relies on AWS and Azure cloud services to scale processing without compromising latency, and on AI agents that automate analysis workflows. The ability to explain decisions through texture-based importance maps (such as GLCM energy or LBP entropy) allows specialists to validate each segmentation, something that transcends mere numerical performance. At Q2BSTUDIO, we develop custom applications that incorporate these techniques, ensuring that each business intelligence or visual analysis solution aligns with the client's real needs. The integration of Power BI to visualize uncertainty calibration metrics, along with cybersecurity strategies to protect sensitive data, completes a robust ecosystem. RadiomicNet demonstrates that it is possible to achieve competitive state-of-the-art performance with compact and understandable models, a path we follow in every digital transformation project.