Automation in Detecting and Localizing Diseased Kidneys in MRI Scans using Deep Learning

Autosomal recessive polycystic kidney disease (ARPKD), a rare genetic disorder affecting 1 in 20,000 children, leads to kidney failure by age 15 due to abnormal renal tubules forming fluid-filled cysts. Diagnosing ARPKD in adolescents prematurely is challenging, as they are in the early stages of kidney development. IMPAKT, a current IT database, helps clinicians identify and manage ARPKD patients early. Traditionally, MRI scans are manually analyzed to collect clinical data, a process prone to errors and inconsistencies. This study introduces a deep learning approach using digital image filtering and artificial neural networks to automate ARPKD detection in DICOMs (digital communication standard for MRI scans). The method accurately localizes cysts by applying spatial high-pass filtering and edge detection, overcoming previous manual detection limitations. This technique can integrate with existing medical records software to predict the likelihood of ARPKD and other genetic diseases at its early stages for risk patients, enabling earlier treatment.

Hello! This is a presentation on the Automation in Localizing and Detecting Disease Kidneys in MRI Scans using Deep Learning. My name is Rushil Johal and with the assistance of the University of Pennsylvania Health System or UPenMed and Children’s Hospital of Philadelphia.

Opening the successfully filed USPTO patent application was a tangible symbol of my biggest dreams. My journey began with sleepless nights, manually tracing and scanning MRI silhouettes of inflamed kidneys to gather vital biological data, such as water volume and iron content. This required drawing 10 outlines per MRI scan layer and each kidney having 40 layers. It’s a cumbersome process for physicians conducting genetic disease testing. I pondered: There must be a more efficient way.

Autosomal recessive polycystic kidney disease (ARPKD) is a rare genetic disorder affecting 1 in 20,000 children. It’s caused by fluid-filled cysts in renal tubules, leading to decreased kidney function and eventual failure by mid-40s. Deep Learning (DL) can revolutionize early ARPKD diagnosis and transform the healthcare industry by collecting and analyzing data from MRI scans for various diseases, potentially saving thousands of lives.

I am studying Computational Data Science and have versatile programming experience in R, Java, MATLAB, and Python. Thus, I apply these hands-on skills in scientific research, specifically interning in MRI Radiology at the University of Pennsylvania’s Perelman School of Medicine and Children’s Hospital of Philadelphia in Summer 2022.

There, I evaluated how an image filtering, DL algorithm is controlled to detect and localize early telltale signs of ARPKD in adolescent patients by examining MRI scans, known as DICOMs.

We‘ll cover the current problems with diagnosing ARPKD and its data collection process, how this DL algorithm functions and its limitations, and the algorithm‘s implications and extensions throughout the medical industry.

To first understand why ARPKD is challenging to detect early and its problematic data collection process, we’ll cover its relevance with DL.

Dr. Anwar Padhani, spokesman for the 2016 International Conference on Radiation Medicine, describes DL branching from ML by utilizing neural networks, like a brain neuron, with many interconnected layers to automatically extract patterns from data, enabling high-level abstraction and complex tasks like image recognition.
ARPKD is difficult to diagnose early due to having no cure, no symptoms nor pain, and delayed development of cysts by the time the risk patient typically reaches 12-14 years old, where tissue trauma has already occurred. Furthermore, the symptoms of excess water retention simulate a steep exponential growth between the patient’s age and kidney size, almost indistinguishable from a healthy, linear-size kidney growth as the patient matures.

Current data collection begins by extruding DICOMS from another picture archiving and communications system, PACs, to be mapped into a standard coordinate system. We collect data by free-hand drawing regions of interest (ROIs) as silhouettes of specific tumors within the kidney. According to nephrologist Dr. Charlotte Gimpel’s 2019 PubMed publication, this tedious methodology is time-consuming, susceptible to human error, unable to localize cysts in DICOMs with blurry qualities, and inconsistent among varying DICOM series.

Understanding DL’s relevancy in streamlining ARPKD’s data, we’ll focus on the algorithm’s operation and limitations.

Digital image filtering and artificial neural networks of multiple input and output automate localizing ARPKD in DICOMs. This steerable algorithm applies high radio frequency and edge detection of image pixels to detect abnormal cysts compared to the rest of the kidney structure. The same approach is applied for Snapchat filters detecting your facial features based on lighting. Once multiple DICOMs of the same patient study and date have been uploaded, the algorithm leverages prior pixel data to improve performance for the same repeated set of tasks.

However, the algorithm may initially experience difficulty differentiating ARPKD’s appearance from a healthy kidney in a DICOM or not generalize to different genetic variations of ARPKD. This can lead to misdiagnosis by false positives or negatives for ARPKD. However, Dr. Kouta Ito’s 2021 publication in the National Center for Biotechnology Information’s findings reveals this algorithm can replace the manual method by decreasing medical uncertainty and random error by 18% with an accurate detection success rate of 93%.

After comprehending the algorithm’s method, let’s focus on its implications and extensions in other fields within healthcare.

This algorithm can be implemented in various medical records software, such as Epic Systems, and includes Multi-Modal Fusion, combining information from multiple imaging modalities, such as CT and ultrasound, to improve localization accuracy. Lastly, the method may detect early patient liver and bile duct cancer risks.

For implications: the algorithm minimizes the time required for diagnosis and treatment planning. It also provides quantitative measurements of cyst size, distribution, and progression over time, helping track disease progression and evaluating treatment efficacy in real-time, not possible prior. Thus, we can tailor personalized treatment plans based on each patient’s severity or unique ARPKD characters. However, legal issues can be a barrier, such as using huge amounts of confidential patient data or a risky financial investment, inaccessible to many underfunded healthcare systems nationwide.

Overall, this image filtering, DL method can change the healthcare industry by diagnosing unaware patients at risk, not just in ARPKD, but any disease studied via MRI DICOMS, even if a cure does not yet exist.

We have covered DL‘s importance in solving ARPKD diagnosis and addressing ARPKD‘s concerns, the new methodology and its constraints, and its impact on healthcare.

Evaluate how you may incorporate DL into your everyday life. Allow an algorithm to teach itself and do the complicated, repetitive tasks for you better after each try. After all, ML is just a baby learning how to walk, but sooner than you realize, will drive a car.

Thank you so much for your undivided attention!