Learn to Recognize

Illustrative figure by Shadi Albarqouni

We started investigating Convolutional Neural Networks for Object Recognition in a supervised fashion, for example, mitotic figure detection in histology imaging (Albarqouni et al. 2016), Catheter electrodes detection and depth estimation in Interventional Imaging (Baur et al. 2016), femur fracture detection in radiology (Kazi et al. 2017), in-depth layer X-ray synthesis (Albarqouni et al. 2017), and pose estimation of mobile X-rays (Bui et al. 2017). One of the first work which has been highly recognized and featured in the media is AggNet (Albarqouni et al. 2016) for Mitotic figure detection in Histology Images. Although the network architecture was shallow, it was trained using millions of multi-scale RGB patches of histology images, achieving outstanding performance (ranked 3rd among 15 participants in AMIDA13 challenge).

During our work, we found out such data-driven models demand a massive amount of annotated data, which might not be available in medical imaging and can not be mitigated by simple data augmentation. Besides, we found out such models are so sensitive to domain shift, i.e., different scanner, and methods such as domain adaptation is required. Therefore, we have focused our research directions to develop fully-automated, high accurate solutions that save export labor and efforts, and mitigate the challenges in medical imaging. For example, i) the availability of a few annotated data, ii) low inter-/intra-observers agreement, iii) high-class imbalance, iv) inter-/intra-scanners variability and v) domain shift.

To mitigate the problem of limited annotated data, we developed models that Learn from a Few Examples by i) leveraging the massive amount of unlabeled data via semi-supervised techniques (Baur and Albarqouni et al. 2017), ii) utilizing weakly labeled data, which is way cheaper than densely one (Kazi et al. 2017), iii) generating more examples through modeling the data distribution (Baur et al. 2018), and finally by iv) investigating unsupervised approaches (Baur et al. 2018, Baur et al. 2019).



  • Siemens Healthineers
  • Siemens AG
Shadi Albarqouni
Shadi Albarqouni
Professor of Computational Medical Imaging Research at University of Bonn | AI Young Investigator Group Leader at Helmholtz AI