Peng team
AI FOR MICROSCOPY AND COMPUTATIONAL PATHOLOGY
Tingying Peng - Group leader
AI TO HELP MICROSCOPY IMAGE ANALYSIS
Tingying Peng’s Helmholtz AI young investigator group’s goal to create new AI methods to help life scientists and pathologists to analyze microscopic images more quantitatively and efficiently, allowing them to extract more knowledge. We would like to develop deep learning methods to address several unique challenges for biomedical imaging, including:
- Domain differences between medical images and natural images where most of deep learning techniques originate;
- Scarcity of high-quality annotated data for efficient network training; and
- The need for explainable algorithms, which usually conflicts with the ‘black-box’ nature of deep neural networks.
The key research strategy we proposed is interpretable deep learning, which blends domain knowledge such as conventional model-based methods and deep learning-based algorithms in a ‘maximise a posterior’ (MAP) manner. This combined approach will leverage the advantages of both model types allowing us to make more accurate predictions while also shedding light on the underlying mechanisms of neural networks for making those predictions.
Research lines
- Artificial Intelligence for Microscopy: We focus on developing novel AI-based algorithms for microscopy image processing, including cell segmentation, detection, classification and quantification. Particularly, we target a few key challenges in microscopy images:
- Real biological signal mingled with experiment noise and batch variations.
- Inadequate or no ground-truth labels available.
- Datasets of highly unbalanced classes.
-
We work on both classic microscopy modalities, such as bright-field and fluorescence microscopy, and advanced ones, such as Cryo-electron tomography (Cryo-ET, in collaboration with Dr. Ben Engel’s group at Helmholtz Pioneer Campus) and extended depth-of-field (EDOF) microscope with “Electrically Tunable Lenses” (ETL, in collaboration with Dr. Jan Taucher in GEOMAR)
- Computational Pathology: Our goal is to develop novel AI methods to aid trained professional pathologists in their decision-making, with a special emphasis on three aspects:
- Self-supervised hierarchical image feature extraction that highlights a small number of key pixels leading to accurate diagnoses.
- Confident measures that assess the reliability of AI predictions.
- Human-in-the-loop processes that make AI models more accessible and explainable by facilitating interaction between algorithms and experts.
-
We collaborate closely with PD. Dr. med Melanie Boxberg in TUM pathology on computational pathology.
