News

Professor C.-C. Jay Kuo, Director of MCL, attended the IEEE Conference on Multimedia Exposition (ICME) held in Niagara Falls, Canada, from July 15-19, 2024. Professor Kuo had dual roles in this conference as a Panel co-Chair and a keynote speaker. Professor Kuo gave his keynote on 7/18 (Thursday) on “Toward Interpretable and Sustainable AI via Green Learning.” Besides, Dr. Kuo and Dr. Zicheng Liu of AMD organized a panel as summarized below.

Panel Title: Generative AI – Opportunities, Challenges, and Open Questions

Panel Background: Generative AI has received a lot of attention due to the tremendous success of ChapGPT. Large foundation models have been trained, leading to various demos and potential applications such as text-to-image and text-to-video cross-domain generations. Resources have been invested in building massive computational and storage infrastructures. Furthermore, data collection and cleaning are essential to high system performance. In the face of these rapid developments, this panel will discuss opportunities, challenges, and open questions associated with generative AI.

Four panelists were invited:

Rogerio Feris, IBM Research

Lijuan Wang, Microsoft Research

Jiebo Luo, University of Rochester

Junsong Yuan, State University of New York at Buffalo

Q&A topics:

Today’s generative AI is tilted more toward “engineering” than “science.” Will this be a concern in the long run?

What are the major shortcomings of the current large foundation models?

How vital are “data collection and cleaning” tasks in generative AI? How do large companies carry out such tasks? Will we run out of data? If so, how soon?

Will “copyright,” “plagiarism,” and “hallucination” be issues? How can we address them? How can we trust the answers?

What roles can small AI companies and academia with limited resources play?

What is the future R&D direction of generative AI? What will be the next big breakthroughs?

Dr. Kuo also had a [...]

Magnetic resonance imaging (MRI) is a good way to detect clinically significant prostate cancer and guide biopsies, due to the superior resolution and contrast of imaging, without harming the human body. Based on prostate MRI, prostate segmentation is a process to localize prostate boundaries for radiotherapy and automate the calculation of the prostate volume. Automatic prostate segmentation is an important step in computer-aided diagnosis of prostate cancer and treatment planning [1].

It is very hard to collect and obtain large annotated datasets for AI In Healthcare. We worked with USC Keck Medical School on this project, and they provided us with a large medical dataset, which was very precious and helpful. In addition to this dataset, we also used some public datasets like ISBI-2013 [2] and PROMISE-12[3] to analyze and evaluate our Green U-Shaped Learning (GUSL) methodology.

Our Green U-Shaped Learning (GUSL) framework is a feed-forward encoder-decoder system based on successive subspace learning (SSL), and it consists of two modules: 1) encoder: fine to coarse unsupervised representation learning with cascaded VoxelHop units, and 2) decoder: coarse to fine segmentation prediction with voxel-wise regression and local error correction. Our model is lightweight and totally transparent while keeping comparable performance.

We have done 5 cross-validations for the dataset from USC Keck Medical School. For T2-cube MRIs, the Dice Similarity Coefficient (DSC) of the prostate segmentation was over 93%. The USC Keck Medical School doctors were very satisfied with these impressive results. In the next step, we will apply our GUSL model to some public datasets and then compare and analyze the performance of our method with some state-of-the-art Deep Learning methods. In the future, we aim to develop methods for segmenting other organs like cardiac. I hope our methods [...]

Professor C.-C. Jay Kuo, Director of MCL, visited Singapore from June 17-23 and attended two events in his Asia trip.For the first event, Professor Kuo gave a seminar titled “Demystify Artificial Intelligence and Technology Outlook” at the Nanyang Technological University on June 20 (Wednesday). The abstract of his talk is given below.“The term “Artificial Intelligence (AI)” was coined in 1956. Although the field evolved slowly in the first 55 years, we have witnessed rapid advances in AI in the last decade (e.g., the ChatGPT service). Questions are raised about AI’s role in human society and civilization, e.g., whether AI will replace human intelligence (HI), how HI and AI complement each other, etc. I will shed light on them. Then, I will comment on the limitations of today’s deep-learning-based (DL) AI models. DL-based models are neither interpretable nor sustainable. An alternative methodology is desired. To this end, I have investigated a new statistically based AI/ML framework called “Green Learning” (GL). GL significantly reduces the model size and complexity of DL models while yielding competitive performance and allowing mathematical transparency. GL adopts the feedforward one-pass training pipeline, so all intermediate results are explainable. Finally, I will address the AI role in human society and issues with today’s chatbots, such as bias and fairness.”Professor Kuo’s talk was hosted by Professor Woon Seng Gan at NTU. He also had a reunion with Professor Jiaying Liu of Peking University. Professor Liu was a visiting PhD student at MCL from August 2007 to August 2008.For the 2nd activity, Professor Kuo delivered a keynote speech at the Workshop titled “Emerging Trends and Innovations in Machine Learning and AI” on June 21 (Friday) organized by APSIPA Singapore Chapter and quite a few [...]

Nuclei segmentation is a fundamental task required to analyze the underlying nuclei structure of an organ of interest. Cancer starts from the cells, and understanding the nuclei shapes, sizes and distribution can provide cues on whether or not a patient has cancer. Further analysis can also help in cancer grading and prognosis. However, studying whole slide images of biopsied tissues requires a large amount of time and effort. 

Such monotonous and laborious tasks can be simplified by using AI and ML, and can perhaps improve the accuracy of the detections as well. Some challenges in the nuclei segmentation task include inherent staining variations in the WSI, a wide variety of shapes and sizes in nuclei, and irregular boundaries which make it difficult to track the actual contours. Most of the current research in this area involves deep learning based architectures like the U-Net, R-CNN, and even Vision Transformer. These methods require a large number of training samples and high complexity to achieve generalization among the variations inherent in nuclei from different organs.

We propose to use a light-weight, interpretable, and simple Green Learning based approach to perform Nuclei Segmentation. Prior work on highly effective Unsupervised Nuclei Instance Segmentation (HUNIS) [1] forms the first stage of our current approach. To further improve HUNIS results, we now focus on the regions where HUNIS requires the help of labels. We divide our current task into two stages: (i) to identify those areas where we need help and (ii) to correct those areas towards their actual class. With the help of Saab Transform, our main task now is to perform feature engineering to identify the ideal features to implement the above two stages. 

References:

[1] V. Magoulianitis, Y. Yang, and C.-C. J. [...]

Surface reconstruction from point cloud scans plays a pivotal role in 3D vision and graphics, with diverse applications in areas such as AR/VR games, heritage preservation, and indoor/outdoor scene reconstruction. This task is inherently challenging due to the ill-posed nature of reconstructing continuous surfaces from discrete points. Furthermore, real-world point cloud scans introduce several obstacles, such as varying densities, sensor noise, and missing parts. These properties make the problem a long-standing one, continually driving researchers to seek more effective solutions.

Early research focused on constructing watertight objects using combinatorial methods [1][2], which inferred the connectivity between points directly. The mainstream of surface reconstruction adopts an implicit surface approach [3][4], where the surface is represented as an unknown continuous function solved by associated partial differential equations (PDEs). Although these methods offer good quality, they are constrained to predicting watertight objects and cannot handle highly distorted LiDAR scenes. Recently, indoor/outdoor scene reconstruction has gained more attention, with deep learning (DL) models [5] demonstrating success in solving this problem based on a supervised learning framework.

Despite their high reconstruction quality, DL models face challenges in generalizability and complexity. In scenarios such as point cloud compression, quality assessment, and dynamic point cloud processing, there is a growing need for low-complexity, low-latency surface reconstruction methods. However, existing DL-based methods often sacrifice simplicity for high reconstruction quality, leaving a gap for low-complexity solutions. We aim to develop an unsupervised few-shot learning method to achieve reconstruction for scenes with low complexity.

Building on our previous unsupervised framework (GPSR), we propose an enhanced version to handle non-watertight indoor/outdoor scenes, named Green Point Cloud Surface Reconstruction++ (GPSR++). The main idea involves building an unsigned distance field (UDF) through approximated heat diffusion and optimizing the surface using [...]