AI-powered treatment planning

Total 16건 1 페이지
Fast CT image reconstruction using GPU

The current trend of CT imaging requires ‘as low dose as possible’. In order to follow this trend, the necessity of fast and low dose image reconstruction algorithm is being emphasized. The main objective of fast reconstruction algorithm is the acceleration of process via optimization of current reconstruction algorithm and utilizing Graphic Process Unit (GPU). For this, an optimization process using a new program language is necessary.

Deep Learning-based Metal Artifact Reduction

Patients selected for cardiac radio-ablation are often associated with implantable cardioverter-defibrillator (ICD)s containing leads connected to pulse generator to stimulate cardiac tissues and prevent sudden cardiac death. These lead implants cause artifacts on Cone Beam Computed Tomography (CBCT) scans that severely hinders the alignment of the target. The purpose of this study is to develop an effective reconstruction algorithm that minimizes the streaking artifact induced by mobile leads of ICDs in the CBCT scan.

Breast Automatic Segmentation

Breast cancer is the most commonly occurring type of cancer among women in the majority of countries, accounting for 1 in 4 newly diagnosed cancer. Breast cancer is the most common site treated with radiation therapy (RT) in Korea, accounting for almost 30% of total patients who underwent RT. 

The precise delineation of the clinical target volume (CTV) and organ-at-risk (OAR) have become important in the era of three-dimensional conformal RT and intensity-modulated RT in breast cancer. The OAR and CTVs in RT planning are currently manually segmented by radiation oncologists. However, given the number and complexity of the structures, the delineation is laborious and often time-consuming.

So we have developed a breast auto segmentation model using deep learning to solve the problems of manually segmentation, which is laborious and time-consuming.

Pelvis GTV Automatic Segmentation

Pelvic cancer is one of the most common cancers in female cancers. For some stages of cervical cancer, the preferred treatment is radiation alone of surgery followed by radiation. Most of the tumors are hard to find in the CT. So most of the OAR (Organs at risk) segmentation is based on CT. However GTV (Gross tumor volume) segmentation is based on MRI.

Contouring GTV in RT planning are currently manually segmented by radiation oncologists. However, given the number and complexity of the structures, the delineation is laborious and often time-consuming.

So we have developed a pelvis GTV auto segmentation model using deep learning to solve the problems of manually segmentation, which is laborious and time-consuming.

Deep Learning based detection and segmentation

Steerotactic Radio Surgery (SRS) and Steerotactic Ablative Radiotherapy (SABR) are type of radiation therapy in which a few very high doses (About 10~20 Gy) of radiation are delivered to small size tumors once or several times. If the radiation is incorrectly irradiated, it is a fetal threat to the patient’s life and requires accurate contouring of the tumor and surrounding OARs. In addition, the advent of MR-Linac and MR-guided radiation therapy tends toward MR-based radiation treatment planning. Accordingly, contouring of tumors and surrounding OARs is required in MR images.

DL based segmentation for breast and H&N cancer

Advancements of radiotherapy developments for breast cancer may render more precise and conformal radiotherapy, designed to target the tumor and avoid normal organs on a much more individualized level. Manual segmentation is the gold-standard in the current radiotherapy planning which, however, is the most time-consuming process and prone to inter-observer variation. Given this issue, interest for auto-segmentation has been rising.

Clinical validation is the process of interpreting the clinical significance of the technology-based research outcomes in an area where medicine and technology coexist. Through in-depth analysis of test patients, we identify the patterns between the anatomical characteristics of each patient and the performance of the model, in order to evaluate the role of the technology in question in a clinical scenario. The results of clinical validation can further help advance the technology by allowing limitations to be identified and feeding them back to the model learning process. 

The aim of our study is to determine the clinical feasibility of auto-segmentation methods for target and normal organs and specifically evaluate the feasibility of a deep-learning-based approach compared to the commercially released atlas-based segmentation solutions.

Clinical application of auto segmentation

In accordance to technical development in radiation therapy, a planning process becomes complicated and important. After planning CT acquisition, segmentation for target area (gross tumor volume, clinical target volume, and planning target volume) and organ-at-risk (OAR), which is called contouring, is needed to identify the 3-dimensional volumetric information for assessing the quality of planning and standardization of planning. Despite several consensus guidelines for contouring, both inter-observer and intra-observer variability remains the issue for standardization and qualified planning. Therefore, an auto-segmentation becomes essential for standardization and efficiency in treatment planning. 

In our research laboratory, we are conducting research on auto-segmentation methods and their clinical implementation.

AI-based image generation

To verify its effectiveness after anticancer treatment, criteria for assessing tumor reactions to treatments were presented, and some of standards are being used with constant development and improvement. However, while evaluation criteria have been revised steadily with an empirical and statistical approach, methods of measuring the condition of a tumor is relatively classic and usually depends on people or tools.

The most widely used criteria as of 2020 is RECIST1.1. After ‘WHO Criteria’ was first published in 1981, followed by Response Evaluation Criteria in Solid Tumors (RECIST) by research groups from Europe and North America in 2000. Finally, after some revised, RECIST1.1 was announced in 2009. It is mainly used for the purpose of assessing the effectiveness of surgery, chemotherapy, and radiation therapy. 

The purpose of this study is to evaluate the treatment response of solid cancers more objectively, accurately and quickly than conventional methods by measuring and classifying the tumor’s condition using deep learning based on RECIST1.1. Furthermore, the ultimate goal is to improve existing assessment criteria and present new standards to provide optimal treatment for each patients.

Generate contrast CT images from Non-Contrast CT images

Recently, there has been a growing interest in breast cancer among cancer species that may develop cardiac toxicity after radiation treatment. The increase in the long-term survivor of breast cancer has shown that the cause of non-cancer death is heart disease caused by dose to the heart during breast cancer radiation treatment. The effect of radiation on the heart should be minimize when receiving breast cancer treatment and using contrast CT in planning radiation therapy can be beneficial to minimizing the dose. However, patients who are unable to take contrast CT due to clinical or other reasons may have difficulty minimizing the dose. Therefore, we want to solve the problem by applying deep learning and creating a contrast CT image only with non-contrast CT images.

Progressive Deep Learning for Segmentation

The recent advancement of Deep Learning (DL) has achieved state-of-art in a wide range of applications especially in object recognition, classification, and segmentation of medical images.

However, training modern deep learning models requires a large amount of computation and long training time due to the complex nature of network structures and a large number of training datasets. Moreover, it is a highly manual and repetitive process to search the optimized configurations of hyper parameters that fit best for the DL network.

So we presented a novel approach to accelerate the training time of DL via progressive feeding of training datasets based on the similarity measures for medical image segmentation. We term this approach as Progressive Deep Learning (PDL).

AI-based 3D prediction for automatic planning

Intensity-modulated radiation therapy (IMRT) is widely used in external beam radiation therapy, however, the planning process requires iterative revision, varies with the planner’s experience, and becomes unintuitive as the complexity of plan objective increases.

The final goal of this study is to automatically predict the optimal treatment dose from CT images and organs-at-risk information only, and further to generate the final treatment plan without repetitive tasks or additional information by the treatment plan optimization process based on the predicted dose.

This study is an intermediate study aimed at developing artificial intelligence-based 3D dose prediction models and predicting more accurate treatment doses using a number of previously conducted retrospective treatment plan information.

Portable Respiratory Training System

In the case of lung and liver cancer patients, the location of cancer cells changes by an average of 2-3 cm and up to 6-8 cm due to breathing. If existing treatments is performed in a state where the cancer cells are not fixed, the uncertainty of the treatment increases, so to solve this problem, the latest treatment methods such as respiratory based 4DRT(4-Dimensional Radiotherapy) are implemented.

Currently, the respiratory interlocking system used when performing 4DRT has limitations in the target patients due to the problems of price, complexity, independent space, and auxiliary personnel. Therefore, this study aims to develop a Wi-Fi-based portable patch sensor breathing training system that can solve the problem of limiting the application target of the existing respiratory linkage system and maximize the therapeutic effect

3D Radiochromic polymer dosimetry

With the advent of new complex and precise radiotherapy techniques such as IMRT, SRS, and VMAT, the demand for an accurate and feasible three dimensional (3-D) dosimetry system has increased. However, the dose gradient in advanced radiation therapy can cause high-dose radiation exposure of normal tissues even with small errors, and may lead to side effects. Therefore, a 3D dosimeter is required, which can measure the radiation dose in the complex radiation therapy. The dose characteristics of the dosimeter were evaluated by improving the sensitivity of a commercial dosimeter.

We investigated the influence of incorporating tartrazine on the dose response characteristics of radiochromic 3D dosimeters based on polyurethane resin. We use three types of polyurethane resins with different Shore hardness values: 30 A, 50 A, and 80 D. Tartrazine (Yellow No. 5) helps incorporate a yellow dye to fabricate the dosimeter. Three sets of six different PRESAGE dosimeters were fabricated to investigate the effects of incorporating yellow dye on the dose response characteristics of the dosimeter. The dose response curve was obtained by measuring the optical absorbance using a spectrometer and optical density using optical CT, respectively. For the optical density measurement, significant sensitivity enhancements of 36.6% and 32.7% were achieved in polyurethane having a high Shore hardness of 80 D and 50 A by incorporating tartrazine, respectively. 

Flexible film dosimeter

When tumors are located on the skin or at superficial regions near the skin, the prescribed doses cannot be delivered to the whole volume of the tumors with high-energy photon beams owing to electronic disequilibrium. Therefore, to increase doses deposited to the patient’s superficial regions, layers capable of increasing the electron fluence, such as boluses, wax, or three-dimensional printed devices, should be placed on the skin.

When applying boluses to patients’ irregular surfaces, such as scalp, breast, perineum, and foot, there might be discrepancies in the setup of the boluses with respect to the treatment plans and the actual deliveries of the plans, which is undesirable. In such cases, in vivo dosimetry is performed to verify accurate delivery of the treatment plans to patients. For in vivo dosimetry, various dosimeters, such as a thermoluminescent dosimeter (TLD), optically stimulated lumi-nescent dosimeter (OSLD), metal oxide semiconductor field-effect transistor (MOSFET), or GAFCHROMIC EBT3 radiochromic films (EBT3), are currently used in the clinical setting. However, the TLD, OSLD, and MOSFET are capable of only point dose measurement, EBT3 is not flexible enough to apply to irregular surfaces. The aims of this study were to develop a flexible film dosimeter applicable to the irregular surface of a patient for in vivo dosimetry and to evaluate the device’s dosimetric characteristics.

A flexible film dosimeter with active layers consisting of radiochromic-sensitive films and flexible silicone materials was constructed. The dose-response, sensitivity, scanning orientation depen-dence, energy dependence, and dose rate dependence of the flexible film dosimeter were tested. The red channel demonstrated the highest sensitivity among all channels, which were the same as the characteristics of GAFCHROMIC EBT3 radiochromic films. The flexible film dosimeter showed no significant energy dependence for photon beams of 6 MV, 6 MV flattening filter-free (FFF), 10 MV, and15 MV. 

Development of Silicon Carbide (SiC) Semiconductor Sensor: for Small field dosimetry and Radiation Hardness

Development of high-resolution and high-strength sensors based on silicon carbide(SiC) semiconductors for small irradiated-high energy radiation measurements

Steerotactic Radio Surgery (SRS) and Steerotactic Ablative Radiotherapy (SABR) are type of radiation therapy in which a few very high doses (About 10~20 Gy) of radiation are delivered to small size tumors once or several times. Accurate dose verification of small field and high doses is very important because radiation is a fatal threat to the patient's life.

Ion detectors have the disadvantage of poor resolution when measuring doses on small fields. Diamond detectors and edge detectors are expensive and have limitations for measuring point doses. In addition, silicon detectors are prone to radiation defects, so we are developing high-intensity semiconductor radiation sensors that can compensate for these problems.

Development of Scintillation Fibre Detector: Real-time Dosimetry

High dose rate (HDR) brachytherapy is a radiotherapy technique that treats cancer by delivering doses using sealed radioactive sources inside the human body. HDR brachytherapy is used to treat cancers of the prostate, rectum, cervix, esophagus, and other organs. 

The radioactive source is moved to nearby tumors with afterloader unit and delivers the radiation dose to treat the cancer. So, the accuracy of source position and occupancy time determines the patient safety and quality of treatment.

We have developed a scintillation fiber-based measurement system capable of real-time verification of brachytherapy and conducted feasibility studies on it.