Interpretable Deep Learning for Fast Medical Image Reconstruction and Analysis


Öksüz İ.(Yürütücü), ÇUKUR T., ÜNAY D., KARAARSLAN E.

TÜBİTAK Projesi, 2020 - 2023

  • Proje Türü: TÜBİTAK Projesi
  • Başlama Tarihi: Şubat 2020
  • Bitiş Tarihi: Şubat 2023

Proje Özeti

The current approach to medical imaging is essentially serial: image acquisition is followed by image analysis and clinical interpretation. In addition, motion is currently resulting in long scanning times for medical images with only a small fraction of the data being used for image reconstruction. This leads to breath-holds that are difficult to tolerate by sick patients. There is an urgent need for precise and reliable Computer-Aided medical Diagnostic (CAD) systems in clinical practice. Unprecedented accuracy achieved by Deep Learning techniques in disease diagnosis is certainly inspiring in this matter. General Deep Learning-based CAD systems treat diagnosis as a classification problem where expert annotated data is used for training the model and accuracy is measured by the Deep Learning model’s performance on previously unseen examples. Deep learning methods for classifying medical images have demonstrated impressive accuracy in a wide range of tasks, but often the process by which images are classified is hard to interpret. These issues are recently exposed in multiple “failure-related” computer vision literature - just carving open the lid of Deep Learning black box in the process. This opacity limits the applicability of these methods in clinical practice.


This Tübitak 2232 proposal, “Interpretable Deep Learning for Fast Medical Image Reconstruction and Analysis” prevents this unfortunate repercussion of Deep Learning-based CAD by mimicking the clinical expert to reduce the reasoning-gap – leading to an accurate and clinically reliable glass box end- to-end CAD system. We focus on cardiac MRI and CT for the following reasons - a virtually incurable disease with growing number of victims indicates urgent need of reliable and accurate CAD system, whereas availability of large annotated public database with strict acquisition protocol ensures a tractable problem for Deep Learning reasoning experiments. However, most of the findings from this study will be generalizable to other diagnostic scenarios. The interpretability will be achieved through modular exploration of clinical and biomechanical constraints, saliency comparison between clinical experts and Deep Learning models as indirect measure of reasoning-gap, and incorporating reasoning-gap feedback for generalization. This project builds on current works on a transformative approach in which acquisition, analysis and interpretation are tightly coupled, with feedback between the different stages in order to optimize the overall objective: Extracting clinically useful information. Developing such an integrated approach to imaging will enable rapid, continuous and comprehensive imaging that is both simpler and more efficient than current practice, eliminating “dead time” between separate specialized acquisitions and allowing extraction of multiple dynamic as well as tissue contrast parameters simultaneously. One approach to address this issue is the use of robust image reconstruction techniques to improve the image quality. The appropriate algorithms can enable shorter scan times without compromising image quality. The current paradigms of compressed sensing and Deep Learning methods offer great potential to improve the image quality of the reconstructed images. The main objective of the proposed project is developing Deep Learning-based algorithms to accelerate the current clinical practice in cardiac imaging to enable faster image acquisition with high quality, which can reduce the health care costs dramatically. The successful completion of this project will lead to a glass-box Deep Learning-based CAD for cardiac diseases that at the same time will be highly accurate, reliable and clinically interpretable.