AI for breast cancer diagnosis

Among the papers presented was a study entitled:

“Improving Diagnostic Confidence in Breast Cancer Detection through Spatial Attention and Uncertainty Modelling.”

Led by a team of students from aivancity under the supervision of Professor Anuradha Kar, this project explores how certain advanced AI techniques, notably spatial attention mechanisms and uncertainty modelling, can improve the reliability of classification models in mammography.

In particular, the proposed approach makes it possible to identify cases where the AI’s prediction is reliable and those requiring validation by a radiologist, thereby strengthening confidence in diagnostic support systems.

Given that breast cancer remains one of the leading causes of death among women worldwide, this research illustrates the potential of artificial intelligence to improve early detection and support healthcare professionals.

This project was carried out by a team comprising:

• Grace-Esther Dong
• Léana A. Yemene Manfouho
• Gil-allen M. Mounzeo
• Johyce D. Bagnenda

On this occasion, Grace-Esther Dong also received a student grant awarded by the conference, in recognition of the quality and scientific potential of the work presented.

A second research project on frugal AI for medical imaging

A second research project from aivancity was also accepted and presented at the conference:

“Domain-Adaptive Auto-Supervision: An Eco-Efficient Pipeline for the Detection of Thoracic Anomalies.”

This work focuses on the automated interpretation of chest X-rays (CXR), a major challenge for the large-scale screening of numerous lung conditions.

Whilst the most powerful artificial intelligence models often require significant computational resources, this research proposes an approach inspired by Green AI, aiming to develop models that are lighter, faster and more data-efficient.

The methodology is based in particular on:

• self-supervised pre-training, enabling the model to learn normal anatomical structures from healthy images
• supervised fine-tuning optimised to detect thoracic abnormalities, particularly rare conditions When evaluated on the NIH ChestX-ray14 benchmark, the method achieves an average AUC of 0.8217, whilst requiring far fewer training iterations and computational resources than conventional approaches.

The research team behind this project comprises:

• Clément Frerebeau
• Marc Habib
• Sélim Jomaa
• Eve Jouni
• Bryan Fozame
• Doreid Ammar
• Anuradha Kar

Training a new generation of AI talent

Through these projects, aivancity actively encourages student research, supporting its students in taking part in scientific conferences and backing their academic initiatives.

This work also illustrates the school’s educational ambition: to train AIgineers® capable of developing artificial intelligence technologies whilst understanding their human, societal and ethical implications.

In a technology sector still marked by gaps in representation, the successes of Grace-Esther Dong and Léana Yemene Manfouho also serve as a reminder of the importance of supporting and promoting the role of women in artificial intelligence and technology.

For the AI of tomorrow can only be responsible, inclusive and beneficial to society if those who design it reflect the diversity of the real world.