LEMMA

Advancing AI for medical imaging

Over the past decade, deep learning has transformed the analysis of medical images, with AI systems now achieving remarkable performance for many diagnostic tasks.

However, these advances come with a major limitation: deep learning for medical imaging depends heavily on vast amounts of labeled data. Creating such data is slow and costly, as it requires medical experts to provide detailed annotations, in some cases down to the level of pixels.

The LEMMA project aims to overcome this bottleneck by developing self-supervised learning (SSL) methods for medical imaging. Unlike traditional approaches, SSL can learn directly from the data, reducing the reliance on manual labeling. By pretraining model parameters this way, SSL offers a path toward more efficient and scalable medical AI models.

LEMMA’s ambitious goals

LEMMA’s goal is to improve the imaging analysis when labels are scarce or noisy, and to make it easier to extend AI tools to new types of medical images. Its approach will also better capture pathological regions, learn from repeated scans of the same patient over time, and identify the most effective way to fine-tune pretrained models for specific tasks.

LEMMA will focus on explainability: developing methods to visualize how self-supervised models learn, extending explainability to image-text models, and designing architectures that are interpretable by design rather than being black boxes. The project will also address the problem of bias in medical datasets, bias which often arises from imbalances or variations between hospitals, scanners, or patient groups. This will make the models more robust when deployed in real-world environments.

The LEMMA project will result in several demonstrators and applications that showcase the practical value of the innovations. For instance, new algorithms will improve the alignment of 3D dental scans with 2D X-rays, enabling better diagnosis and treatment planning. The project will also deliver more accurate tumor segmentation in PET-CT scans of resected tissue samples, robust dental pathology detection from noisy X-ray annotations, and advanced methods for measuring brain atrophy that separate disease effects from acquisition discrepancies.

Finally, the consortium will examine the value of medical foundation models, designing amongst others models capable of executing an array of downstream tasks, including segmentation, landmark annotation, object detection, and classification on CT scans.

Towards faster tool development

By improving generalization and label-efficiency, LEMMA’s advances are expected to outperform current approaches that are based solely on supervised learning. They may also shorten AI development cycles by around 20%, accelerating the release of new tools for digital pathology. This will enable the project partners to release new commercial AI applications faster and to extend the applicability of existing ones.

“By reducing the need for costly manual annotation and making AI more adaptable, LEMMA is expected to cut the development time of AI models, helping medical companies bring new tools to market faster. For doctors, this means more reliable decision support; for patients, it means earlier and more accurate diagnoses, and ultimately better care.”