Dr. Erickson's Radiology Informatics Lab at Mayo Clinic works to develop CT, MRI and ultrasound tools for identifying and classifying biomarkers. Investigators hope these tools can help fine-tune patient treatments options.

  • Adrenal masses. The purpose of this project is to identify imaging-based biomarkers that predict whether an incidentally discovered adrenal mass has a high likelihood of being malignant, thus directing surgical intervention.
  • Body composition analysis. Most CT and MRI examinations of the abdomen are performed to assess the condition of abdominal organs. However, there is an increasing awareness that the composition of one's body has significant value as a predictor of health and the ability to withstand significant stresses such as major surgery. In this project, images of the abdomen are used to compute specific markers of body composition: intra- and extra-abdominal fat, and the thickness of the abdominal musculature. We then assess the value of these as predictors of health and survival. We are also examining related biomarkers for their value as health predictors.
  • Brain lesions classification. The aim of this project is to build a deep-learning classifier to discriminate among different classes of tumefactive brain lesions using MRI: gliomas, metastases, lymphomas and tumefactive multiple sclerosis lesions.
  • Generative adversarial networks to synthetize brain MRI. Missing brain MRIs represent a major obstacle for the development of new deep-learning models and the application of already existing ones. For this project, we aim to synthetize missing brain MRI sequences using generative adversarial networks (GANs) to use as substitutes of real MRIs. To understand if these images could one day represent a valid alternative to real MRIs, we will assess the quality of the generated images using them as inputs of existing deep-learning models. The same approach could be extended to different imaging modalities.
  • Thyroid ultrasound evaluation of nodules. Thyroid nodules are common findings in the population. Differentiating benign from malignant nodules can be a challenge; in addition, not all malignant nodules need to be aggressively treated. This project attempts to identify properties of nodules as imaged by ultrasound for their ability to predict aggressive versus nonaggressive nature.
  • Classification of cardiac diseases. This project uses deep-learning methods applied to cardiac ultrasound imaging (echocardiography) to classify cardiac diseases.
  • Glioblastoma progression and pseudoprogression biomarkers. This project develops a biomarker using early magnetic resonance imaging (MRI) to determine if early enhancement is due to tumor progression or treatment effects that mimic tumor progression. An accurate determination of true versus pseudoprogression is critical — effective therapy should continue during treatment, but if the tumor is progressing, second-line agents can be beneficial.
  • Polycystic kidney disease biomarkers. This project develops efficient, clinically viable tools that allow estimation of prognosis and therapy planning. The efficient characterization of disease can allow for better assessment of disease state and measurement of the benefits of possible therapies.
  • CT brain segmentation and stroke detection. Ischemic stroke is one of the leading causes of cardiovascular morbidity and mortality. Although CT plays an essential role in the evaluation of patients with suspected stroke, it is challenging for radiologists to identify ischemic regions from noncontrast CT alone. In this project, we aim to develop an algorithm that automatically identifies and localizes ischemic stroke regions using both segmentation and classification models.
  • Predicting the dislocation risk following total hip arthroplasty. Total hip arthroplasty (THA) is one of the most successful surgical procedures, as it brings significant pain relief and increased quality of life for patients. Dislocation is a common complication following THA, representing a challenging problem for both patients and surgeons. We aim to develop a deep-learning model to predict the dislocation risk of THA patients based on their postoperative radiographs.
  • Measurement of acetabula component angles following total hip arthroplasty. The risk of dislocation following THA is related to several patient-related, implant-related and surgical-related factors. Acetabular component mispositioning is one of the most important and modifiable risk factors for postoperative instability and dislocation. Acetabular inclination (abduction) and anteversion angles define the orientation of the acetabular component. We aim to develop deep-learning segmentation models and image processing algorithms that can fully automate the measurement of the acetabular component angles on postoperative anterior-posterior or cross-table lateral radiographs.
  • Fast MR imaging reconstruction — super resolution. MRI, as a noninvasive medical-imaging modality, is widely used and powerful for clinical diagnosis and research, yet often limited by its slow acquisition speed. There is an active demand in practice to reduce the acquisition time, in order to relieve the burden of patients, to reduce examination cost and decrease the motion-induced artifacts. We aim to develop novel deep-learning models to achieve state-of-the-art quality acquisitions through the both intraslice and interslice orientations, which can be potentially applied to the clinical diagnosis and radiographic study.
  • DL-based decision support tool for distinguishing uterine leiomyosarcomas. Despite recent advances in imaging, it is still challenging to differentiate uterine leiomyosarcomas (LMSs), a rare but aggressive cancer, from leiomyomas (LMs), a highly prevalent benign condition. Recently artificial intelligence and deep-learning techniques have increasingly been utilized in various areas including imaging diagnosis. We hypothesize that using cutting-edge texture analysis, machine learning and general artificial intelligence-based approaches can revolutionize the role radiological imaging plays in diagnosis of uterine tumors.
  • Artificial intelligence (AI) to auto-segment median nerve and distinguish CTS using ultrasound imaging. Carpal tunnel syndrome (CTS) is characterized by tingling and numbness in the distribution area of the median nerve due to compression of the median nerve at the carpal tunnel. A necessary precursor is a method to accurately and reliably measure nerve morphology in situ in the clinical setting. Our central hypothesis is that changes in median nerve morphology, such as volume and shape, reflect underlying changes in nerve function and physiology, which in turn are useful predictors of treatment outcome. For this purpose, ultrasound is in many ways the ideal method, since it is noninvasive, painless, inexpensive and portable. AI tracked and segmented median nerve statistics could give insight on CTS symptoms and assist radiologists in decision-making.
  • DL to fully automate segmentation of polycystic kidneys and compare performance in MRI versus US imaging. Deep-learning techniques are becoming the leading algorithmic approaches to solve inherently difficult image-processing tasks. MR imaging examinations are widely used for kidneys of patients affected by polycystic kidney disease (PKD). Ultrasound (US) could be an alternative with advantage being portable, easily available across globe, quick and inexpensive. We hypothesize that deep-learning techniques can be leveraged to perform fast, accurate, reproducible and fully automated segmentation of polycystic kidneys. Further performance on PKD classification will be tested by both imaging techniques (MRI versus US) and determine if the US could be an alternative or supportive approach for PKD patients.