UoL participated at the IEEE ISBI’24 Conference

The IEEE International Symposium of Biomedical Imaging (ISBI) is the premier international venue for progress in processing of biological imaging and the currently fast developing field of histopathology imaging. It is also one of the most impactful conferences for the fields of biomedical and radiological imaging. This year the ISBI conference was held in Athens, Greece, at the Megaron Athens International Conference Center, from May 27^th to 30^th. The conference received 50% more paper manuscript submissions this year compared to previous years and was eventually attended by 1,200 participants. This reflects the fast developments in the field due to progress in imaging modalities, Artificial Intelligence (AI)-based methodologies, as well as principled based methodologies and their applications to the fields related to the conference. The event in Athens was productive and successful and enabled open discussion about current progress in the field and future directions.

The conference held a panel discussion on challenges and perspectives of translating AI research in biomedical imaging into clinical practice. A member of the panel was Constantinos Daskalakis, MIT. Concerning verification, he pointed out that even a complicated AI system may produce an output that is simple to verify and gave as example the protein folding tool. With respect to generative AI he pointed out that to model the complexity of high-dimensional data the size of the training data would have to increase exponentially which is not feasible. He argued for the need to model only a limited number of statistics for a small number of features. One example is assumptions made for MRI reconstruction. He gave as main example the genome, where the size of a complete model of the genome would have to be “out of this world,” whereas modeling SNPs (Single Nucleotide Polymorphisms) is feasible.

In terms of interpretability of AI, Constantinos said that it is hard to achieve for deep learning systems that have billions of parameters. It is however possible to understand which part of the input causes the output. For example, it is possible to understand which corpus of text produces the generated text. Due to the limitations of interpretability, he suggested that we should use AI systems for applications that benefit everybody and for tedious tasks without taking the final decision. For general AI, he addressed the overhype and bold statements for the potential of AI made by companies. He said academia is supposed to be more measured and objective. Companies are good at scaling systems. Academia looks at a system at a higher level for a more general evaluation. He does not believe we are close to general AI or to a doom’s day scenario; however, we should be careful so that AI does not fail in bad way. When asked to provide three appropriate adjectives for AI he said, “complementarity to humans,” “challenging,” and “promising. “

In his plenary talk, Joseph Sifakis, Director at Verimag laboratory of CNRS, France, he said that we currently only have weak AI for specific objectives, such as intelligent question answering assistants. He believes that further progress can be achieved by advancing the relationship between computational processes and the physical world. In particular, relate AI to autonomous systems. An example of such systems is the Internet of Things (IoT) and more generally ICT.

There was also a plenary talk by Anant Madabhushi, Emory University, on the evaluation of AI in clinical trials. He discussed some of his group’s ongoing work in validating AI algorithms in the context of radiology and pathology for precision medicine. He gave several examples of both completed and prospective ongoing clinical trials.

Another plenary talk was by Francis Bach, Inria and Ecole Normale Supérieure, France. He talked about diffusion models that have impressive results in generative AI. His talk proposed a simpler formulation for the models that does not involve stochastic differential equations.  Despite the progress of generative AI, he said that generative AI cannot give rise to new data not represented in the training data set.

In general, the methodology that was emphasized at the conference was on established as well as modern techniques of deep learning-based AI for biological and biomedical image analysis. Some of these are the transformers to represent long range dependences and context in an image. Another is the use of diffusion models for generative AI that can also be used as reconstruction prior instead of the sparsity prior, for example.  There were several sessions on the topics of deep learning and generative AI. Elaborations of AI presented included methods and tools for explainability of deep learning.

There were multiple poster and oral sessions for various aspects of microscopy image analysis for biological and biomedical microscopy. They were for imaging data of bacteria, cells, cell cultures, as well as histology and histopathology. Data from multiple microscopy modalities were included such as fluorescence, light-field, and electron microscopy. The imaging data were stationary, volumetric, and even video. Announcements for open databases for benchmarking of methodologies were also combined with results from international competitions. The methods developed were for detection, segmentation, registration, cell motion tracking, and tissue characterization.

Many sessions were on radiological imaging. They were mostly for MRI of the brain. But problems from other modalities were included such as reconstruction of CT data, as well as processing of EEG or EKG data. The processing of this data was with statistical or machine learning methodology.

Stathis Hadjidemetriou from the Department of Information Technologies of University of Limassol was a co-chair for the oral Session on “Signal and image reconstruction.” During this session four papers were presented on MRI and CT data, primarily for improved priors for reconstruction based on deep learning diffusion methods. These methods can also be applied to microscopy.

Stathis presented two posters at the conference. One poster was for a work that is the result of the collaboration of the Department of Information Technologies of University of Limassol and the Department of Biology of the University of Cyprus. It was for their latest research on cellular optical microscopy imaging data analysis. This groundbreaking work focuses on automated cell division quantification from phase contrast microscopy, significantly boosting throughput in assessing novel anticancer drugs. The method developed first identifies potentially dividing cells in individual time frames, especially with the DeepLabV3+ deep learning network. The specific network is able to identify cells having a variety of sizes and even shapes. The method then streamlines the process along time to achieve a high performance. This research is part of the ABiOMiCeC project, supported by the Cyprus Research and Innovation Foundation from the RESTART 2016-2020 Programmes, co-funded by national and European funds.

The second poster presented by Stathis was for the results of a project in collaboration with the radiological institute at SHK Nordhausen, Germany on Whole-Body (WB) MRI. How would it be if one could undergo a whole-body cancer screening in a single MRI session? It sounds like science fiction, but we are close to achieving this objective. The radiological institute at SHK Nordhausen, under the direction of Prof. Ansgar Malich, following current scientific trends offers the so-called ‘WB MRI’ screening as an outpatient service. In 40-minute sessions of free breathing, the MRI explores all the tricky corners of the human body and detects cancerous lesions down to 1 cm with 100% success. This exam has already been tested in a series of approximately 1,000 patients. A further recent achievement of the institute at SHK of making WB-MRI safe and trustworthy without the use of contrast enhancer (Gadolinium) was published in an internationally acknowledged journal.

In collaboration with our international partners at SHK Nordhausen, and the Deutsches Krebsforschungszentrum (DKFZ) in Heidelberg, Germany, we continue to pioneer by emphasizing on the applicability of whole-body magnetic resonance imaging (WB-MRI). We have now shifted our focus to post-acquisition computational improvements of the acquired MRI data to improve image quality and enhance overall efficacy. To this end, Stathis, in cooperation with Dr. Papageorgiou from the Institute for Radiology in SHK Nordhausen, have developed and implemented a software based on ‘computer vision’ principles. This software removes abrupt ‘shading artifacts’ from the multiple consecutive coils along the patient. The restored images provide a biologically and anatomically significant image that allows for faster and more reliable interpretation.

Whole-body magnetic resonance imaging (WB MRI) is a guideline-approved method for radiation-free detection of bone metastasis. However, thanks to the work of Stathis and other scientific efforts worldwide, WB MRI will soon become a competitive tool for cancer screening and follow-up of all body organs, including the liver, kidneys, spleen, and adrenals.

We were excited to share our findings with the international community at the ISBI’24 conference and engage in discussions that enhance our perspectives on methodology, its validation, and on providing the best clinical care.

The paper on WB MRI will be made available from IEEE Xplore.

The ABiOMiCeC project acknowledges the support of the Cyprus Research and Innovation Foundation within the context of the RESTART 2016-2020 Programmes, which are co-funded by national and European funds.