Skip to main content.
kumc.edu
Reliable Intelligence & Medical Innovation Lab

Research

Below are current research projects underway in the Reliable Intelligence and Medical Innovation Laboratory (RIMI Lab) at the University of Kansas Medical Center.

Choose a research project to learn more:

The goal of reliable AI is to build a unified AI model that can achieve balance, credibility, adaptation, and interpretation simultaneously. To achieve this goal, multi-objective and brain-inspired learning will be developed for model training. Different from the state-of-art training-testing patter, model adaptation stage will be introduced to minimize the discrepancy between trained model and testing sample. To obtain reliable output, the credibility of model output will be measured as well. Then the final decision can made accordingly.

A flowchart representing the stages of AI model development. Starting with 'Initial Model' followed by 'Multi-objective training' and 'Brain-inspired training,' leading to a 'Trained Model.' Then it moves to 'Model Adaptation,' followed by 'Model Testing,' which leads to two outcomes: 'Predictive Output' and 'Credibility Measurement,' culminating in 'Final Decision.'

The principal challenge in building reliable artificial intelligence (RAI) is uncertainty, while probabilistic inference is an ideal solution. Evidential reasoning (ER) is a general probabilistic inference engine, which is potential to build RAI model that can achieve robust, safe, balance and interpretation in a unified way.

Mathematical equation defining 'p_h' as 'k' times the ratio of two complex products. The numerator involves products of 'w_i p_{h,i}' over '1 plus w_i minus r_i' across 'i' from 1 to 'N', subtracted by products of '1 minus r_i' over '1 plus w_i minus r_i'. The denominator is '1 minus k' times the product of '1 minus r_i' over '1 plus w_i minus r_i', across 'i' from 1 to 'N'. The variable 'h' ranges from 1 to 'H'."

ER rule (ER2) is a new development of ER, which introduces both reliability and weight into the probabilistic inference. However, the original ER2 is a recursive algorithm, which can not perform inference in a straightforward way. As such, we inferred the analytic ER2 in theory. The inferred analytic formular is shown as follows:

Mathematical equation defining 'p_h' as 'k' times the ratio of two complex products. The numerator involves products of 'w_i p_{h,i}' over '1 plus w_i minus r_i' across 'i' from 1 to 'N', subtracted by products of '1 minus r_i' over '1 plus w_i minus r_i'. The denominator is '1 minus k' times the product of '1 minus r_i' over '1 plus w_i minus r_i', across 'i' from 1 to 'N'. The variable 'h' ranges from 1 to 'H'."

An illustration of evidential reasoning rule learning pipeline. A flow diagram illustrating a process for determining the reliability and uncertainty estimation of a test sample. On the left, an oval labeled 'Test sample' points to a vertical green rectangle representing an input layer. Inside this layer, there are three boxes aligned vertically, labeled 'Prior Probability,' 'Likelihood probability Feature 1' with 'r1' in red, 'Likelihood probability Feature 2' with 'r2' in red, and continuing down to 'Likelihood probability Feature M' with 'rm' in red, indicating multiple features being considered. Dotted lines suggest additional features between Feature 2 and Feature M. To the right of this layer is a wider green shape labeled 'AER' for 'Adaptive Evidence Reasoner,' with 'Reliability l' above it, signifying a processing step. This shape points to two ovals on the right side: the top one is labeled 'Probabilistic Output,' which leads with a green arrow to another oval labeled 'Label.' Below it is an oval labeled 'Overall Reliability,' which also leads with a green arrow to another oval labeled 'Uncertainty Estimation.' This diagram shows a clear progression from a test sample through processing to produce a label and an estimation of uncertainty.

Then we are developing a new evidential reasoning rule (ER2) learning strategy to overcome these challenges. In ER2, the trained model will be tuned for each test sample by introducing individual reliability and the reliability of model output can be evaluated through overall reliability. Several ER2 based ML/DL algorithms are developed.

We also developed belief rule based (BRB) model for clinical diagnostic support by integrating domain knowledge into model construction. However, the model safety and robustness have not been considered yet. We are developing a new reliable rule base which integrates safe, robustness and interpretation into a unified model and ER2 is used to perform inference. To improve model performance further, the model parameter/architecture optimization algorithms are investigated as well.

"Text describing a rule-based system with a rule Rk. If x1 is Ak1, x2 is Ak2, through xMk is AkMk, then the output y is a series of decisions D1, β1,k, to DN, βN,k with a constraint that the sum of βn,k from n equals 1 to N must be less than or equal to 1. Includes rule weight θk, attribute weights δk1 through δkMk, and reliability for each test sample ri."

Featured publications:

  1. Z. Wang, Q. Wang, J. Wu, M. Ma, Z. Pei, Y. Sun, Z. Zhou*, “An Ensemble Belief Rule Base Model for Pathologic Complete Response Prediction in Gastric Cancer”, Expert Systems With Applications, Vol 233, Issue 15, 2023
  2. Jie Wu, Qianwen Wang, Zhilong Wang, Zhiguo Zhou*, “AutoBRB: An automated belief rule base model for pathologic complete response prediction in gastric cancer”, Computers in Biology and Medicine, 140, 105104, 2022
  3. Zhiguo Zhou, Kai Wang, Michael Folkert, Hui Liu, Steve Jiang, David Sher, Jing Wang, “Multifaceted radiomics for distant metastasis prediction in head & neck cancer”, Physics in Medicine and Biology, Vol. 65, 155009, 2020 (Featured by Physics World Magazine)

Developing a balanced, robust and safe model is desirable in many real-world applications, particular in medicine. We are developing automated multi-objective learning (AutoMO) to achieve this goal. AutoMO is new ML/DL training-testing pattern. Instead of training single model in training stage, AutoMO produces a Pareto-optimal model set by optimizing multiple objectives simultaneously to build the balanced model. In testing stage, multiple Pareto-optimal models are selected and fused through ER adaptively for different test sample to achieve robustness. Meanwhile, the overall reliability is obtained to measure the safety of output. The hyperparameter optimization is introduced to tune the hyperparameters automatically.

An illustration of AutoMO workflow

Featured publications:

  1. Zhiguo Zhou, Liyuan Chen, Michael Dohopolski, David Sher and Jing Wang, “ARMO:  automated and reliable multi-objective model for lymph node metastasis prediction in head and neck cancer”, Physics in Medicine and Biology, 68, 095012, 2023

  2. Xi Chen, Jiahuan Lv, Dehua Feng, Xuanqin Mou, Ling Bai, Shu Zhang, Zhiguo Zhou*, “AutoMO-Mixer: An automated multi-objective Mixer model for balanced, safe and robust prediction in medicine”, International Workshop on Machine Learning in Medical Imaging (MLMI 2022)

  3. Zhiguo Zhou, Meijuan Zhou, Zhilong Wang, Xi Chen, “Predicting treatment outcome in metastatic melanoma through automated multi-objective model with hyperparameter optimization”, SPIE Medical Imaging Conference, 2022 (Oral Presentation)
  4. Xi Chen, Meijuan Zhou, Zhilong Wang, Si Lu, Shaojie Chang, Zhiguo Zhou, “Immunotherapy treatment outcome prediction in metastatic melanoma through automated multi-objective delta-radiomics model”, Computers in Biology and Medicine, 138, 104916, 2021
  5. Zhiguo Zhou, Michael Dohopolski, Liyuan Chen, Xi Chen, Steve Jiang, David Sher, Jing Wang, “Reliable lymph node metastasis prediction in head & neck cancer through automated multi-objective model”, IEEE International Conference on Biomedical and Health Informatics (BHI), 2019

Although deep learning has reached remarkable success, there exist fundamental difference between artificial neural networks’ operating procedures, particularly concerning continuous or lifelong learning processing. We are investigating the brain-inspired learning mechanisms and applying them into neural network model training.

Through analyzing a large number of quantitative image feature, radiomics has achieved great success to improve personalized treatment assessment, cancer screening and treatment outcome prediction. However, there are several challenges that limit its generalizability and reliability. As such, we are developing a new multifaceted radiomics (M-radiomics) framework It consists of four stages: 1) Cross-institution image acquisition and regeneration; 2) Prior knowledge-based image segmentation; 3) Hierarchical feature extraction & learning; 4) Multi-attribute predictive model construction.  Several strategies for each stage have been developed.

Figure 1: Multifaceted radiomics pipeline. Four stages of a new multifaceted radiomics (M-radiomics) framework: 1) Cross-institution image acquisition and regeneration; 2) Prior knowledge-based image segmentation; 3) Hierarchical feature extraction & learning; 4) Multi-attribute predictive model construction.  

To capture the difference between pre- and post- treatment images, delta radiomics was developed. We have developed automated multi-objective delta radiomics model to consider the balance between sensitivity and specificity. However, the biggest challenge is to achieve the consistence between pre- and post- treatment images. We are developing new coherent delta (Co-delta) radiomics model to overcome this challenge.

Figure 2: An illustration of delta radiomics study. Training process and training process

We also investigated radiogenomics which is aiming to discover the connection between image and gene information. A multi-classifier multi-objective (MCMO) model was developed to predict gene mutation through radiomic features. We are also working on the multi-omics study by integrating multiple omics information to build more comprehensive model.

Featured publications:

  1. Xi Chen, Meijuan Zhou, Zhilong Wang, Si Lu, Shaojie Chang, Zhiguo Zhou*, “Immunotherapy treatment outcome prediction in metastatic melanoma through automated multi-objective delta-radiomics model”, Computers in Biology and Medicine, 138, 104916, 2021
  2. Zhiguo Zhou, Kai Wang, Michael Folkert, Hui Liu, Steve Jiang, David Sher, Jing Wang, “Multifaceted radiomics for distant metastasis prediction in head & neck cancer”, Physics in Medicine and Biology, Vol. 65, 155009, 2020
  3. Zhiguo Zhou, Shulong Li, Genggeng Qin, Michael Folkert, Steve Jiang, Jing Wang, “Multi-objective based radiomic feature selection for lesion malignancy classification”, IEEE Journal of Biomedical and Health Informatics, 24 (1), 194-204, 2020
  4. Kai Wang, Zhiguo Zhou, Rongfang Wang, Liyuan Chen, Qiongwen Zhang, David Sher, Jing Wang, “A multi-objective radiomics model for the prediction of locoregional recurrence in head and neck squamous cell cancers”, Medical Physics, Vol 47, Issue 10, 5392-5400, 2020
  5. Xi Chen, Zhiguo Zhou*, Raquibul Hannan, Kimberly Thomas, Ivan Pedrosa, Payal Kapur, James Brugarolas, Xuanqin Mou and Jing Wang*, “Reliably Predicting Gene Mutation in Clear Cell Renal Cell Carcinoma through Multi-classifier Multi-objective Radiogenomics Model”, Physics in Medicine and Biology, 63 (21), 2018

Accurately predicting treatment outcome or response in cancer therapy can help physicians make better and more personalized treatment plan so as to improve patient’s survival time and quality. Our goal is to develop a reliable treatment outcome prediction model that unifies balance, safe, robustness and interpretation into one framework and translate it into clinical practice. Currently, we are investigating two problems in cancer therapy: (1) Treatment follow-up prediction in multiple cancer sites, including distant failure/metastasis, lymph node metastasis, locoregional recurrence predictions. (2) Immunotherapy response prediction in metastatic melanoma, lung cancer, head & neck, etc.

A visualized definition of treatment outcome prediction in cancer therapy

Featured publications:

  1. Xi Chen, Meijuan Zhou, Zhilong Wang, Si Lu, Shaojie Chang, Zhiguo Zhou*, “Immunotherapy treatment outcome prediction in metastatic melanoma through automated multi-objective delta-radiomics model”, Computers in Biology and Medicine, 138, 104916, 2021
  2. Zhiguo Zhou, Kai Wang, Michael Folkert, Hui Liu, Steve Jiang, David Sher, Jing Wang, “Multifaceted radiomics for distant metastasis prediction in head & neck cancer”, Physics in Medicine and Biology, Vol. 65, 155009, 2020 (Featured by Physics World Magazine)
  3. Zhi-long Wang, Li-li Mao, Zhiguo Zhou, Lu Si, Hai-tao Zhu, Xi Chen, Mei-juan Zhou, Ying-shi Sun, Jun Guo, “Pilot study of CT-based radiomics model for early evaluation of response to immunotherapy in patients with metastatic melanoma”, Frontiers in Oncology, 10, 1524, 2020
  4. Kai Wang, Zhiguo Zhou, Rongfang Wang, Liyuan Chen, Qiongwen Zhang, David Sher, Jing Wang, “A multi-objective radiomics model for the prediction of locoregional recurrence in head and neck squamous cell cancers”, Medical Physics, Vol 47, Issue 10, 5392-5400, 2020
  5. Zhiguo Zhou, Genevieve M. Maquilan, Kimberly Thomas, Jason Wachsmann, Jing Wang, Michael R. Folkert, Kevin Albuquerque, “Quantitative PET Imaging and Clinical Parameters as Predictive Factors for Patients with Cervical Carcinoma: Implications of a Prediction Model Generated Using Multi-Objective Support Vector Machine Learning”, Technology in Cancer Research & Treatment, Vol 19: 1-9, 2020
  6. Rongfang Wang, Yaochung Weng, Zhiguo Zhou, Liyuan Chen, Hongxia Hao, Jing Wang, “Multi-objective ensemble deep learning using electronic health records to predict outcomes after lung cancer radiotherapy”, Physics in Medicine and Biology, 64, 24, 245005, 2019
  7. Liyuan Chen, Zhiguo Zhou, David Sher, Qiongwen Zhang, Jennifer Shah, Nhat-Long Pham, Steve Jiang, Jing Wang, “Combining Many-objective Radiomics and 3-dimensional Convolutional Neural Network through Evidential Reasoning to Predict Lymph Node Metastasis in Head and Neck Cancer”, Physics in Medicine and Biology, 64 (7), 2019 (Featured by Physics World Magazine)
  8. S. Li, B. Li, Z. Zhou, N. Yang, H. Hao, M. Folkert, P. Iyengar, K. Westover, H. Choy, R. Timmerman, S. Jiang, and J. Wang, “A pilot study using kernelled support tensor machine for distant failure prediction in lung SBRT”, Medical Image Analysis, 50, 106-116, 2018
  9. Zhiguo Zhou, Michael Folkert, Puneeth Iyengar, Kenneth Westover, Yuanyuan Zhang, Hak Choy, Robert Timmerman, Steve Jiang, Jing Wang, “Multi-objective radiomics model for predicting distant failure in lung SBRT”, Physics in Medicine and Biology, 62, 4460-4478, 2017
  10. Z. Zhou, M. Folkert, N. Cannon, P. Iyengar, K. Westover, H. Choy, R. Timmerman, S. Jiang, and J. Wang, “Predicting distant failure in early stage NSCLC treated with SBRT using clinical parameters”, Radiotherapy & Oncology, 119 (3), 501-504, 2016

Clinical diagnostic support can significantly impact improvements in quality, safety, efficiency, and effectiveness of health care. It plays an important role in cancer screening, staging, malignancy prediction and other diseases. Our aim is to develop a trustworthy diagnostic support system and translate it into clinical practice. Two strategies have been developed, they are data driven based model through machine learning/deep learning and knowledge driven based model through rule base. These methods have been applied to malignancy prediction in breast cancer, lung nodule classification, lymph node metastasis diagnosis, etc.

path of diagnostic support model illustrated. Figure 1: Malignancy lesion detection in breast cancer through digital breast tomosynthesis

path of diagnostic support model equation. Figure 2: A cooperative belief rule base model for lymph node metastasis prediction in gastric cancer

Featured publications:

  1. Chen, X. Wang, J. Lv, G. Qin and Z. Zhou*, “An integrated network based on 2D/3D feature correlations for benign-malignant tumor classification and uncertainty estimation in digital breast tomosynthesis”, Physics in Medicine and Biology, 68, 175046, 2023
  2. Zhiguo Zhou, Shulong Li, Genggeng Qin, Michael Folkert, Steve Jiang, Jing Wang, “Multi-objective based radiomic feature selection for lesion malignancy classification”, IEEE Journal of Biomedical and Health Informatics, 24 (1), 194-204, 2020
  3. Zhiguo Zhou, Genggeng Qin, Pingkun Yan, Hongxia Hao, Steve Jiang, Jing Wang, “A shell and kernel descriptor based joint deep learning model for predicting breast lesion malignancy”, SPIE Medical Imaging Conference, 2019
  4. Shulong Li, Panpan Xu, Bin Li, Liyuan Chen, Zhiguo Zhou, Hongxia Hao, Yingying Duan, Michael Folkert, Jianhua Ma, Shiying Huang, Steve Jiang, Jing Wang, “Predicting lung nodule malignancies by combining deep convolutional neural network and handcrafted features”, Physics in Medicine and Biology, 64, 17, 175012, 2019
  5. Benjuang Yang, Yingjiang Wu, Zhiguo Zhou, Shulong Li, Genggeng Qin, Liyuan Chen, Jing Wang, “A collection input based support tensor machine for lesion malignancy classification in digital breast tomosynthesis”, Physics in Medicine and Biology, 64, 23, 235007, 2019
  6. Zhi-Guo Zhou, Fang Liu, Ling-Ling Li, Li-Cheng Jiao, Zhi-Jie Zhou, Jian-Bo Yang, Zhi-Long Wang, “A cooperative belief rule based decision support system for lymph node metastasis diagnosis in gastric cancer”, Knowledge-based systems, 9 (85), 62-70, 2015
  7. Zhi-Guo Zhou, Fang Liu, Li-Cheng Jiao, Zhi-Jie Zhou, Mao-Guo Gong, Xiao-Peng Zhang, “A bi-level belief rule based decision support system for diagnosis of lymph node metastasis in gastric cancer”, Knowledge-based systems, 54, 128-136, 2013
  8. Zhi-Guo Zhou, Fang Liu, Li-Cheng Jiao, Zhi-Long Wang, Xiao-Peng Zhang, Xiao-Dong Wang, Xiao-Zhuo Luo, “An evidential reasoning based model for diagnosis of lymph node metastasis in gastric cancer”, BMC Medical Informatics and Decision Making, 13 (123), 2013

Medical image processing plays more and more important role in modern medicine, which includes acquisition, reconstruction, enhancement, analysis, visualization and management.

We have developed several knowledge-based methods for tumor segmentation in multiple sites. However, most of current methods only integrates prior knowledge and don’t consider posterior knowledge. Our goal is to develop new delta segmentation method which considers both prior and posterior knowledge in the segmentation.

diagram of dataset to image pre-processing, test/train split, training, post-processing to segmented image. An illustration of tumor segmentation in head & neck cancer

We are also working on image regeneration development so that the images data produced from different protocols can be regenerated in the same standard. This is the fundamental pre-task to build large scale image based prediction model.  

Featured publications:

    1. Ahmad Qasem, Hui Liu, and Zhiguo Zhou*, “Automated tumor localization and segmentation through hybrid neural network”, SPIE Medical Imaging Conference, 2022
    2. Liyuan Chen, Chenyang Shen, Zhiguo Zhou, Kevin Albuquerque, Michael Folkert, Jing Wang, “Automatic PET cervican tumor segmentation by combining deep learning and anatomic prior”, Physics in Medicine and Biology, 64, 8, 085019, 2019
    3. X. Liang, L. Chen, D. Nguyen, Z. Zhou, X. Gu, M. Yang, J. Wang, S. Jiang, “Generating Synthesized Computed Tomography (CT) from Cone-Beam Computed Tomography (CBCT) using CycleGAN for Adaptive Radiation Therapy”, Physics in Medicine and Biology, 64, 12, 125002, 2019 (“Roberts Best Paper Prize”)
    4. Liyuan Chen, Chengyang Shen, Zhiguo Zhou, Genevieve Maquilan, Kimberly Thomas, Michael R. Folkert, Kevin Albuquerque, Jing Wang, “Accurate segmenting cervical tumor in PET based on similarity between adjacent slices”, Computers in Biology and Medicine, 97 (6), 30-36, 2018
    5. Zhi-Guo Zhou, Fang Liu, Li-Cheng Jiao, Xiao-Dong Wang, Shui-Ping Gou, Shuang Wang, “Object information based interactive segmentation for fatty tissue extraction”, Computers in Biology and Medicine, 43 (10), 1462-1470, 2013

The extracellular deposition of amyloid-beta (Aβ) plaques stands out as a defining characteristic of Alzheimer’s disease (AD), a progressively debilitating neurodegenerative condition. Accurate quantification of these plaques proves essential for early diagnosis, tracking disease progression, and deepening our comprehension of the underlying pathological mechanisms. We are aiming to develop deep automated multi-objective (Deep-AutoMO) model and multi-task deep learning (ML-DL) model to characterize the Aβ plaques accurately and reliably in a fully quantitative way.

KU School of Medicine

University of Kansas Medical Center
Department of Biostatistics and Data Science
3901 Rainbow Boulevard
Mailstop 1026
Kansas City, KS 66160