Total 10건 1 페이지
Basic AI/ML research for Trustworthy Medical AI system
Reliable Medical AI Systems to apply in Clinical Practice
  • Various research results have been published that apply artificial intelligence technology to medical problems such as automatic of medical images in limited purposes.
  • However, many researchers point to the development of a reliable medical artificial intelligence model as a major prerequisite for applying AI technology to real clinic.
  • AI as an auxiliary tool for clinical decision-making can help the medical staff to make decisions for diagnosis or treatment of patients only by securing the reliability of the results.
  • The methodologies/topics to deal with the technical limitations of AI models that are obstacles to securing such reliability:
    1. Explainable/Interpretable AI
    2. (Self) Attention Model/Causal Learning
    3. Uncertainty Quantification for Model Robustness
    4. Fairness
Standardization of AI-based Medical Devices and Medical Big Data
Standard Evaluation Technologies for Quality Advancement of Clinical Data
  • For effective use of medical big data, reliability and interoperability of medical data must be secured.
  • However, research on reliability and compatibility of data is limited, due to the diversity of mechanical performance and processing techniques of measuring devices.
  • Based on measurement standards, our research team intend to establish a standardized full-cycle infrastructure or platform that reflects the entire medical workflow from data acquisition, measurement, evaluation, storage, and utilization and to establish high-quality medical big data-based technology by evaluating its effectiveness.
Standard Evaluation Technologies for Quality Advancement of Clinical Data
  • Development of international standards and reference data for data labeling and performance evaluation methods for various AI-based medical devices for diagnosis/prediction of diseases.
    1. International standard development of AI-based medical device (AI-MD) performance evaluation method
    2. Development of international data standards for the performance evaluation of AI-MD
    3. Development of reference data for performance evaluation of AI-MD
    4. Implementation and verification of labeling tool SW module supporting data standards
Development of low energy X-ray intensity variable radiation therapy technique for pets

In the United States, about 4 million pets are diagnosed with cancer and this is 2.5 times higher than that of men being diagnosed with cancer. The treatment method is of importance, however, there are very limited treatment methods compared to the human. With 100 years of history, radiation therapy is seen as one of the most effective treatment methods for various cancer not only for human but for pets as well. In contrast to human cancer treatment, which has been in the field for 100 years, cancer treatment machine optimized for pets does not exist. Due to the expensive cost of radiation therapy machine, few numbers of large animal hospitals abroad use radiation therapy machines designed for humans to treat pets.

Development of kilovoltage X-ray Monte Carlo simulation and 3D printed compensator

In this research, we are developing the system for intensity modulation radiotherapy using kilovlotage X-ray and compensator. 

The Monte Carlo Simulation using GEANT4 was developed to understand the characteristics of kV X-ray irradiator and to calculate the delivered dose into the treatment region.

The simulation include the modeling of “XRAD320”, about Commercial kV X-ray irradiator.

The PDD and profile was measured by the phantom(PTW RW3) and ion chamger(PTW Farmer) and compared with the result from simulation.

Also, a compensator to modulate the intensity of radiation was designed from  treatment region of CT DICOM image and was printed using 3D printer.

Radiation Treatment Planning system for companion animal

Recently, interest in pet health care has been increasing as the number of pets has increased rapidly. As the aging of companion animals progresses, the proportion of companion animals diagnosed with cancer is 2.5 times higher than that of humans. The necessity and demand for companion animal cancer treatment is increasing, but a radiation cancer treatment device optimized for animals has not been developed. In the absence of cancer treatment equipment for companion animals, it treats companion animals with high-energy radiation therapy equipment for the human body and because of the high cost of introducing and maintaining equipment, radiation treatment is possible only in a few large animal hospitals. Since most pets are smaller in size than humans and do not have a deep tumor location, it is expected that cancer treatment will be possible with low-energy radiation. Therefore, this study aims to develop a low-energy radiation therapy device for companion animals and an AI-based radiation therapy planning system.

Metal Artifact Reduction (MAR) in CT

Our work focuses on developing a fast MAR method based on a constrained beam-hardening estimator representing the underestimated error between the expected and calculated reconstruction images. The proposed estimator is derived from a polychromatic X-ray attenuation model with respect to the X-ray transmission length, avoiding dependencies regarding the X-ray spectrum and material attenuation coefficients. It maximizes the accuracy for the correction of beam-hardening artifacts by analyzing the change in the attenuation coefficient level while polychromatic X-ray passes through a homogeneous metallic material. The entire process is completed by a linear combination of two images reconstructed only once, leading to faster computation. The estimator-associated parameters are numerically calculated from an uncorrected CT image and metal-only forward projection. The only unknown parameter to minimize the beam-hardening artifact is fine-tuned by solving linear optimization on the reconstruction image domain without forward and backward projection transformations. This method is sufficiently effective in terms of the optimization speed and quality in MAR, compared with other beam-hardening correction methods.

Comparison among LIMAR, NMAR, BCMAR, and the proposed method for the JawSimulationPhantom (Cu). From left, the uncorrected image, LIMAR result, NMAR result, BCMAR result, and the result of the proposed method are shown. The first row shows the results of each method, and the second row shows the difference images for each result with respect to the ground truth.

Low dose CBCT reconstruction

This study develops an improved Feldkamp–Davis–Kress (FDK) reconstruction algorithm using non-local total variation (NLTV) denoising and a cubic B-spline interpolation-based backprojector to enhance the image quality of low-dose cone-beam computed tomography (CBCT). The NLTV objective function is minimized on all log-transformed projections using the steepest gradient descent optimization with an adaptive control of the step size to augment the difference between a real structure and noise. The NLTV denoising is used to increase the difference between the actual structure and noise of the projection data. The cubic B-spline interpolation is applied to update the voxels in the reconstructed image by resampling 16 adjacent pixels in the projection domain during backprojection.

APICCS based 4D CBCT reconstruction

This study extends APICCS algorithm to 4D CBCT reconstruction, which uses partial projections sorted from projections taken at 3-D CBCT acquisition protocol. The APICCS algorithm is the introduction of the segmentation information classified from prior images in addition to the original prior images. The use of segmentation information makes it possible to construct an adaptive relaxation map and thus adjusts the effects of the updates that would have been made to the unchanged and changed regions of the prior images during the reconstruction process. It improved the image quality of 4-D CBCT and led to a potential imaging dose reduction by using a regular 3-D CBCT scan

Patient-specific Quality Assurance Using a 3D-printed phantom for breast IORT

This study aims to confirm the usefulness of patient-specific quality assurance (PSQA) using three-dimensional (3D)-printed phantoms in ensuring the stability of IORT and the precision of the treatment administered. In this study, patient-specific chest phantoms were fabricated using a 3D printer such that they were dosimetrically equivalent to the chests of actual patients in terms of organ density and shape around the given target, where a spherical applicator was inserted for breast IORT treatment via the INTRABEAMTM system. Models of lungs and soft tissue were fabricated by applying infill ratios corresponding to the mean Hounsfield unit (HU) values calculated from CT scans of the patients. The two models were then assembled into one. A 3D-printed water-equivalent phantom was also fabricated to verify the vendor-provided depth dose curve. Pieces of an EBT3 film were inserted into the 3D-printed customized phantoms to measure the doses.

Unsupervised Machine Learning-Based Approach for Pretreatment Delivery Quality Assurance in CyberKnife System

This study presents an unsupervised machine learning-based pretreatment delivery quality assurance (DQA) for minimizing the beam delivery time using a small number of representative nodes only in CyberKnife system. The spatial node locations and MUs of beams used in patient plan are first extracted from the beam data list. An unsupervised machine learning-based approach, the fuzzy c-means algorithm, is then used to group among the related nodes spatially. The number of group, c, is set to 10 and each beam is arbitrarily assigned to be a cluster. The initial mean value of the 10 clusters is calculated using the spatial node locations of corresponding beams of each group. Depending on exponential weights based on differences between the spatial node locations of each beam and the current means of 10 clusters, each beam is updated and the process is repeated until no change is observed. A representative node in each group is finally determined by selecting the node location of a beam with maximum MU in each cluster.

The original DQA scheme took about 22.7 min from 20 CK DQA plans to complete beam delivery. FCM clustering–based DQA reduced the delivery time to 12.1 min, leading to 46.7% shorter than the original DQA scheme.