Generative AI algorithms development

• Developing GDAL framework for Generative diffusion augmented learning with limited data resources

• Optimizing Gen AI algorithms

• Generative AI algorithms development for adaptation deep learning algorithms

Multi-modal AI Algorithms development

• Multimodal AI models for breast cancer detection

• Multi-modal model development in Geoscience

LLMs/VLLMs

• Language/Vision-Language model in Medical Imaging

• Faithfull Generative Diffusion Vision-Language Model

AI-Computer Vision

Geoscience (collaborative project with Geoscientist Dr Kibria, NSF funded)

• Petrographic/SEM microscope imaging

  • Deep learning algorithms development for fully automated particle segmentation

  • An In-house morphometric algorithms development

Biomedial microscope imaging (collaborative project with MechanoGenomics Lab, Dr Chowdhury)

• Stem cell pattern recognition for advancing cell biology research

Breast cancer detection using AI

• Personalized risk prediction in breast cancer

• Served as a postdoctoral researcher (2020-2024) in "Personalized breast cancer screening using sequential Mammogram" (NIH/NCI R37 project), for developing a personalized breast cancer screening tools using deep learning and time-series mammograms"  under the mentorship of Prof Lee and Prof Nishikawa.  Our research goal, so far for first time, is to develop a near-term mammography-detectable breast cancer risk marker using a new image transformation, deep learning, and statistical approach for longitudinal data on serial mammograms that will help women choose their own personalized protocol in consultation with their doctor. https://reporter.nih.gov/project-details/9940199

• Breast Lesion Detection in DBT

• Developing a 2.5D AI-DL algorithms  for detecting breast lesions in Digital Breast Tomosynthesis (DBT) images

Developed a deep learning-based 2.5D AI model of breast lesion detection in DBT image with Juhun Lee, PhD, and evaluated the method in the AAPM co-sponsored DBT lesion detection Grand Challenge (DBTex, Phase 1). We developed an 2.5D lesion detection algorithm for DBT image (3D-Mammogram) using volumetric morphological operation.  

1. M. B. Hossain, R. M. Nishikawa, and J. Lee, “Developing breast lesion detection algorithms for Digital Breast Tomosynthesis: Leveraging false positive findings,” Medical Physics, 2022, doi: 10.1002/mp.15883.

2. M. B. Hossain, Robert M. Nishikawa, and Juhun Lee "Improving lesion detection algorithm in digital breast tomosynthesis leveraging ensemble cross-validation models with multi-depth levels", Proc. SPIE 12033, Medical Imaging 2022: Computer-Aided Diagnosis, 120330D (4 April 2022); https://doi.org/10.1117/12.2611007

Image recognition 

• Computer-aided Surgical support System using AI

The orthopedic surgery support system aims to improve the quality of surgery by automatically assisting the surgical staff, shortening the operation time and facilitating information sharing among the staff. Therefore, we propose a navigation system that guides surgical staff such as nurses, doctors during surgery, and also can be used for training junior surgeons using image feature extraction, artificial intelligence, and augmented reality. The navigation system consists of three major parts- surgery video acquisition, recognition of surgery scene and  display/sharing result/information among the team members. We used smart eye glasses for hand-fee video recording during surgery. The main challenges of recognition are surgery images are confounded by non-homogeneous illumination, random head motion of operator etc. So, I applied surgical image recognition technology using deep learning to automatically recognize surgical procedures. We are working in the recognition section which is subdivided into- (a) surgery tools detection and (b) workflow recognition in surgery videos. We modified Darknet-53 deep neural network and YOLO3 algorithm for surgery tools detection because it directly can’t be used in surgery case.  Secondly, we utilized phase elapsed time feature from surgery video and proposed CNN-LSTM framework for surgery phase detection. Proposed AI-based surgical navigation system sequentially recognize current surgical procedure and also progress of the surgery through video analysis using deep learning, and then sharing the information in real-time among team members. Thus it can be useful for preparing next procedure in time. We achieved 62.3~96.4% mAP accuracy for tools detection, and 64.4~ 93.1% accuracy for phase recognition 

1. B. Hossain, S. Nishio, H. Takafumio and S. Kobashi, "A deep learning approach for surgical instruments detection in Orthopaedic surgery videos using transfer learning," Proc. SPIE Medical Imaging: Image-Guided Proc., Rob. Interv., & Modeling (Feb. 14, 2020). doi: https://doi.org/10.1117/12.2550670

2. S. Nishio, B. Hossain, M. Nii, T. Yagi, H. Takafumio and S. Kobashi, , "Surgical Phase Recognition Method with a Sequential Consistency for CAOS-AI Navigation System" in IEEE Global Conference on Life Sciences and Technologies, Mar. 2020. doi: https://doi.org/10.1109/LifeTech48969.2020.1570619203

Predictive model in medicine 

• Implanted knee kinematics prediction

This work aimed to predict postoperative knee functions of a new patient prior to total knee arthroplasty (TKA) surgery using machine learning, because such prediction is essential for surgical planning and for patients to better understand the TKA outcome. This study introduced PCA-based generalized linear regression analysis and optimized predictive variables during training the models to predict the most likely surgery outcome, specifically for TKA, before the surgery, by measuring only the preoperative knee functions. The methods were validated for two types of knee functions, and acceptable performance was achieved. Presented in Annual meeting of Orthopaedics Research Society (ORS), Austin, Tx, Feb. 2019

1. B. Hossain, T. Morooka, M. Okuno, M. Nii, S. Yoshiya, and S. Kobashi, "Surgical outcome prediction in total knee arthroplasty using machine learning," J. of Intelligent automation & Soft Computing, Vol. 25, No .1, pp. 105-115, 2019., PDF

2. B. Hossain, and S. Kobashi, "Prediction of Personalized Postoperative Implanted Knee Kinematics with Statistical Temporal Modeling," in Multidisciplinary Computational Anatomy: Springer, Singapore, 2022, pp. 275-281. https://doi.org/10.1007/978-981-16-4325-5_36

Cerebral aneurysm risk prediction

Although lifestyle habits and genetics are considered as the main issues causing cerebral aneurysms, however, according to some studies, the shape of cerebral arteries could be correlated with the risk of aneurysms occurring. This work proposes a method to evaluate the shape of the circle of Willis to predict occurrence of cerebral aneurysms. The experimental results showed that our defined 3-D shape indices of the circle of Willis are good candidates to predict the occurrence of cerebral aneurysms. Furthermore the risk could be influenced by lifestyle habits. Future work will predict the aneurysm occurrence risk by integrating the proposed method and lifestyle practices as well. (Left Fig. ) Anatomical feature of circle of Willi extraction using semi-automated method. (Right Fig.) Model accuracy, The relationship between features F-score and accuracy.)

1. M. Yasugi, B. Hossain, M. Nii, and S. Kobashi, "Relationship between cerebral aneurysm development and cerebral artery shape," J. of Advanced Computational Intelligence and Intelligent Informatics, Vol. 22, No. 2, pp. 249-255, 2018. PDF

2. M. Yasugi, B. Hossain, M. Nii, M. Morimoto, and S. Kobashi, "Cerebral aneurysm occurrence prediction by morphometric analysis of the Willis ring," Proc. of 2016 IEEE Int. Conf. on Systems, Man, Cybernetics (SMC), pp. 1774-1779, 2016. https://doi.org/10.1109/SMC.2016.7844495

Total hip arthroplasty 

Predicting cortical hypertrophy in THA: Quantification of gap between implant and thigh bone for predicting cortical hypertrophy in total hip arthroplasty using X-ray images of hip joint- its consists of segmentation and feature extraction for predictive modelling, - currently working on U-Net based fully-automated segmentation of gap region.

Shape quantification of Knee bone using SSM

This study proposed a method for shape quantification of distal femur in order to substitute subject-specific evaluation. Patient-specific femoral coordinate system (FCS) is essential for biomechanical analysis, patient-specific knee surgery design, and image alignment because the anatomy of the knee differs from ethnicity, even from patient to patient. Furthermore, automated subject-specific FCS determination could reduce time for analysis. Although there are some studies which proposed FCS for lower extremity (whole femur and tibial bone), however, they cannot be directly applied to the distal femur. Therefore, this thesis proposes a fully-automated subject-specific FCS determination method using magnetic resonance (MR) images, and then the defined FCS is applied to align the training subjects to construct statistical shape model (SSM) which is essential to extract morphological shape variation of the distal femur. 

1. B. Hossain, M. Nii, S. Yoshiya, and S. Kobashi, "Fully-automated femoral coordinate system definition for constructing statistical model of distal femur," Int. J. of Biomedical Soft Computing and Human Sciences, Vol. 22, No. 2, pp. 73-83, 2017. PDF

2. B. Hossain, M. Nii, S. Yoshiya, and S. Kobashi, "Computer-Aided Knee Surgery: Automated Anatomical Axis Definition of Clinical Interest," Proc. of 17th International Symposium on Advanced Intelligent Systems (ISIS), Sept 2017 (Best Paper Award).