Research

Lab of Artificial Intelligence in Biomedical Imaging

Medical Imaging

TEE image during cardiopulmonary resuscitation segmentation using deep learning and monogenic filtering

Transesophageal echocardiography (TEE) is an imaging method that uses echocardiography to obtain a high-quality image of the heart and blood vessels. Using this method, continuous TEE images can be obtained during Cardiopulmonary resuscitation (CPR), and the effect of CPR can be quantitatively verified by observing changes in the motion vector and volume of the left and right ventricles during the maximum diastolic and systolic phase. In order to quantitatively calculate the motion vector and volume, it is important to accurately segment the left and right ventricles.

Since the data is very rare, it is possible to use a small number of problems in artificial intelligence by dramatically increasing the number of data by using TPS transformation. Learning was conducted by applying the latest deep learning algorithm. In addition, we improved the segmentation accuracy by applying the concept of monogenic signals as well as deep learning models. So, in this way, we were able to achieve a high level of splitting results.

Central foveal thickness prediction after surgery through deep learning

The treatment of the epiretinal membrane may be followed by observation if there are no symptoms, but if vision decreases or dystrophy occurs, it is removed through vitrectomy. However, no clear indications for surgery have yet been suggested. In 70 to 90% of patients after the epiretinal membrane removal surgery, improvement of visual acuity and dystrophy was observed. However, since there are cases where dysphagia persists after surgery, the discomfort of the patients is not resolved in many cases, and the recurrence rate after surgery is also reported to be as high as 12% to 20.5%. Although several studies have been conducted on the diagnostic value of deep learning, many studies have not yet been conducted on determining whether to treat treatment and predicting the prognosis after diagnosis. Therefore, we intend to develop a tool that can predict postoperative macular thickness using preoperative optical coherence tomography images using deep neural network learning. In addition, further improved results could be obtained by utilizing clinical information.

Kidney segmentation in children ultrasound images using active contour and deep learning model for kidney volume measurement

There is a direct relationship between kidney size and kidney function. Therefore, the change in kidney size becomes an important measure for evaluating the state of the kidney in patients with kidney disease. Therefore, the size change is the most important prognostic factor. In particular, knowing the standard value of the normal kidney size according to the age of the child in the process of growth and accurately measuring the size of the kidney is a great help in diagnosing and predicting the progress of kidney disease. It is important to measure the volume of the kidney and using computed tomography or magnetic resonance imaging can accurately measure the volume. Therefore, we conducted a study to segment the kidney in the coronal section and the transverse section using U-Net. Based on the above result, it will be possible to derive the kidney growth function according to the growth by calculating the kidney volume.

Ultrasound skin tumor image diagnosis using combined deep learning method

Among medical imaging techniques, ultrasound imaging technology is applied in various fields including the dermatology field because it enables rapid examination without using radiation which is harmful to the human body. However, it is easily exposed to noise, and due to the complexity of the human body, ultrasound technologies tend to have a slightly lower diagnostic accuracy. Therefore, in this study, by applying the deep learning technique that occupies SOTA in many diagnostic fields these days, we try to help specialists to improve diagnostic accuracy. In fact, as a result of the study, many meaningful results have been drawn that can help the diagnosis of specialists.

Convolutional neural network for classification of acral melanoma and benign nevi

Melanoma is originated from the melanocyte producing the melanin which determines the complexion, and it has the highest mortality among skin cancers. Among its various subtypes, acral lentiginous melanoma (ALM) occurring in the acral sites are rare and have a relatively higher incidence in non-whites, and is the most common subtype of melanoma in Asian countries. ALM arises from extremities such as hands, feet or fingernails and the shape of ALM at the early stage is similar to benign nevi (BN). Also, it is difficult for non-experts, even for dermatologists, to extract helpful features from dermoscopy, and specialized training is required to obtain improved diagnostic accuracy. Thus, we are researching a computer aided system using a convolutional neural network to improve the accuracy of ALM diagnosis by dermatologists.

Image registration for radiation tracking system accuracy improvement

Radiation therapy is a treatment method that is in the spotlight of patients in that it is a painless and odorless surgery unlike conventional incisional surgery. In order to perform treatment, it is important to identify the precise position of the treatment area by assembling diagnostic information from Radiography. The purpose of this study is to obtain the depth information for the therapeutic radiation through matching using kV X-ray images and Digitally Reconstructed Radiograph (DRR) image obtained from on-board imaging (OBI). To complete the accurate tracking system, a precise registration of kV and DRR images is required. To improve the registration process, we use image enhancement and segmentation techniques.