My research interests focus on exploring how artificial intelligence, specifically computer vision techniques, can be used to enhance patient prognosis in the medical field. I strongly believe that AI can help physicians and healthcare professionals in the near future by accelerating the diagnosis process and offering more personalized medicine. Therefore, any advance in medical image analysis is a step forward in the right direction.

However, despite the rapid growth of the AI field, most of the systems developed for the medical domain remain largely distant from clinical practice. Thus, developing methodologies that bridge the gap between AI and trustworthy clinical applications is crucial.

Data validation and integrity

Data constitute the raw material of AI models. In recent years, a large number of datasets have been published, which has facilitated model training. However, not all studies carry out a thorough understanding of the data, often leading to incorrect conclusions.

In this line of research, we emphasize that data integrity is fundamental. In previous work, I developed an algorithm for detecting “duplicate” images within a dataset, providing a curated and more reliable version known as Curated BUSI.

• C. Aumente-Maestro, J. Díez, B. Remeseiro, A multi-task framework for breast cancer segmentation and classification in ultrasound imaging, Available at Computer Methods and Programs in Biomedicine - Elsevier Journal

Resource-efficient deep learning models

As part of this research line, I developed a lightweight deep learning method and explored image analysis methodologies to enhance the efficiency and interpretability of medical imaging systems in limited computational scenarios for brain tumor segmentation in MRI.

The trend over recent years has been to build more powerful AI systems at the cost of drastically increasing computational requirements (e.g., LLMs, VLMs, VLAs). Finding a balance between achieving sufficiently good performance in delineating brain tumors and managing resource consumption is necessary. From a practical standpoint, this is essential not only due to resource limitations in hospitals but also to support environmental sustainability by promoting green AI and reducing the carbon footprint.

• C. Aumente-Maestro, D. R. González, D. Martinez, B. Remeseiro, BTS U-Net: A data-driven approach to brain tumor segmentation through deep learning, Available at Biomedical Signal Processing and Control - Elsevier Journal

• C. Aumente-Maestro, J. Díez, B. Remeseiro, A multi-task framework for breast cancer segmentation and classification in ultrasound imaging, Available at Computer Methods and Programs in Biomedicine - Elsevier Journal

Evaluation of medical image segmentation AI models

While AI has demonstrated its potential to transform medical imaging workflows, the field still faces ongoing challenges. Despite the efforts to evaluate AI systems, general limitations persist. Most tools primarily focus on evaluating model performance using standardized metrics; however, they often overlook the critical aspect of evaluating model improvements over time. There is also a lack of mechanisms for assessing and visualizing potential shifts between training and test sets (domain shift), and a high demand for sophisticated analysis into model behavior across diverse clinical scenarios, particularly in those cases deemed as “challenging”.

In response to these clinical and technical challenges, in this research line I presented AUDIT, an open-source Python library designed to advance the evaluation of AI models in medical imaging. The main objective here is to develop a unified framework that supports both population-level and subject-level performance analysis, facilitates longitudinal performance assessment, and enables visualization of potential data shifts between training and testing.

• C. Aumente-Maestro, M. Müller, B. Remeseiro, J. Díez, M. Reyes, AUDIT: An open-source Python library for AI model evaluation with use cases in MRI brain tumor segmentation, Available at Computer Methods and Programs in Biomedicine - Elsevier Journal

Research ethics

I strongly believe in:

• Open publishing
• Papers with code
• Diversity in research
• Slowing down the publishing cycle