Week 2 - Hannah Sheffield

 This week, I shadowed Dr. Allison Levy and Dr. Carolyn Eisen for breast imaging. When shadowing Dr. Levy, I saw her diagnose patients based on their mammograms, ultrasound, and MRI images. She taught me some new terminology such as birads, tomosynthesis, and hypoechoic images. With Dr. Eisen, I shadowed a mammogram biopsy where I learned about the necessary procedures used for obtaining a tissue sample. I was intrigued with how I was able to see how tissue samples were obtained, and were given to a pathologist for further analysis. 

As for research, I delve into researching our problem on shifting the delayed arterial input function (AIF - concentration of the contrast agent in the artery throughout time) to align with the reference AIF. To specify, we have a map with a delayed response (called delay map) in capturing the contrast agent concentration, where the TK parameters undergo a loss throughout each epoch loop. To understand the problem on a deeper level, I needed to understand what was happening in the code. I did this by mathematically deriving the gradients of the TK parameters throughout the epoch loops. This gave me a deeper understanding in why a loss occurs between each epoch iteration, and we found that vp (blood plasma volume) is the AIF curve, ktrans (volume transfer constant - rate of contrast agent leaking from the blood vessel to the tissue) is the integral of the AIF, and the delay map is the AIF's derivative (or change in slope) throughout time. Through this, we were able to pin-point the two issues. The first is that the delay map has a very negligible temporal change compared to the TK parameters, which leads to an imbalance between the two metrics. Second is that all variables have very negligible values, which prohibits the epoch loop from capturing the values and iteratively shifting the delayed AIF with the reference AIF. 

To counteract these issues, Ari and I decided to explore options to resolve the issue. First, we artificially changed the learning rate of the delay map to check if there is a response. In changing the learning rate from 1e-2 to 1000, there was an appropriate response where delay map shifted. We then explored normalization techniques to see if the map changes there rather than arbitrarily changing the learning rate. We looked into a gradient normalization technique called "GradNorm," which takes the magnitude of the gradients of the delay learning rate, ktrans, and vp so that the magnitude imbalance is resolved. We found a working code that fixes this issue, and attempted to integrate the GradNorms function into our working code. However, this attempt was unsuccessful (when the learning rate was at its original value), and did not shift the delayed AIF. To this end, we have decided to move forward to the next stage of the project, which is running the working code through all cases, and developing a two compartment model pipeline. We hope to revisit this issue at a later time.

Comments

Post a Comment

Popular posts from this blog

Week One: Eddie Wei

Over the first week of immersion, I had the opportunity to observe a couple standard urologic operating room procedures, each offering a different perspective on diagnosis, imaging, and operative care. On June 1, I observed two standard operating room procedures. The first involved the treatment of four urinary bladder stones using a cystoscope inserted through the male urethra. The cystoscope was first equipped with a specialized head capable of generating a stream of water to break down the stones. After fragmentation, the water head was replaced with a suction tube to remove the stone fragments and excessive water from the water jet. The procedure was performed under direct real time visualization with a  standard 3x magnification camera integrated into the cystoscope. This procedure lasted 2-3 hours, with 2 and a half hours allocated to breaking down these stones using the water jet.  The second procedure involved a female patient undergoing blue-light cystoscopy to evalua...
Read more

Week 1- iqra

This week started off pretty slow since I didn’t meet with my clinician until Wednesday, but things picked up quick ly after that. One thing that has made a big difference is that Dr. Cheung sent an email out to the department about me shadowing, which has allowed me to move around and observe a variety of electrophysiology procedures rather than being limited to a single physician or case type. This has given me exposure to many different aspects of the field and has helped me gain a broader understanding of cardiac electrophysiology. In addition to clinical observations, I also met with one of the medical students working on my research project. During our meeting, we discussed some of the technical details and logistics of the project and worked through a few questions regarding the overall direction of the research. We are still in the early stages of planning, so there are additional details that need to be worked out, and we plan to continue those discussions next week. On Wednes...
Read more

Week 1: Katelyn

This week marked the beginning of our immersion experience. So far, I have had the opportunity to start my research and to shadow my clinician, Dr. Hansen, at his clinic. These appointments were very interesting to observe largely due to the variety of patients that Dr. Hansen has, which include orthopedic oncology, metabolic bone diseases, and trauma cases.  On Tuesday, Dr. Hansen holds appointments with current patients which includes post-ops and follow-ups. I was particularly fascinated by the process of these appointments and Dr. Hansen's ability to quickly analyze the patient's situation. This Tuesday, there were a total of 35 patients including: post-ops for hip and knee replacements; pain follow-ups; osteonecrosis; hip, sacrum, and humerus fractures; osteoporosis treatment follow-ups; osteosarcoma and benign bone tumors; Paget's disease; and repeat stress fractures. Typically, the flow of seeing these patients included a quick review of their file and any scans foll...
Read more