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

ARPKD (MIM #263200) is a rare inherited genetic disorder, affecting 1 in 20,000 children, characterized by the renal tubules of kidneys developing structurally abnormal, resulting in uncontrollable fluid-filled cystic dilations. The nonfunctioning tubules may crush adjacent tubules, resulting in congenital hepatic fibrosis, decreased kidney function, and usually total kidney failure by age 15.

ARPKD is often challenging to diagnose early due to the undevelopment of kidney function and cystic ducts. Regardless, the symptoms for risk patients are kidney volume and water retention in reference to excessive, exponential growth of ARPKD compared to healthy, linear growth as the child matures. Although there is no cure, IMPAKT, a novel IT solution, assists clinicians in identifying and managing patients, who are invention warranted, across hospital systems efficiently and earlier for treatments before ARPKD becomes a life-threatening condition.

MR images are extruded from PACs to be mapped into a standard coordinate system, used to collect clinical organ data (e.g. volume or iron amount) by free-hand drawing region of interests (ROIs) of silhouettes of specific soft tissue or tumors within an organ. In nephrology, this tedious manual process of data collection is time-consuming, susceptible to random error for analyzing each DICOM frame, unable to localize cysts under DICOMs with busy qualities, and inconsistent among varying DICOM series (e.g. COR T2 BLADE FS).

This research uses deep learning, specifically digital image filtering processing, and artificial neural networks of multiple input and output layers to automate localization and detection of ARPKD in DICOMs. This proposed method with steerable features applies spatial highpass, signals with higher frequency than a certain cutoff frequency of various DICOM frames and edge detection in the pixels (dark versus light) of MRI to distinguish abnormal cysts in ARPKD compared to the rest of the kidney structure. Once multiple MRI standardized series of the same patient study and date have been uploaded, the method leverages prior pixel data to improve performance for the same repeated set of tasks. Hence, the software gradually iterates in formulating more accurate ROIs concerning the cysts by inputting thousands of DICOMs.

Although the limitation includes the method initially experiencing difficulty in differentiating appearance of ARPKD to a healthy kidney in a DICOM, findings of this software reveal total replacement of the prior manual method of detection along with decreased medical uncertainty and likelihood of random error by 18% with an accurate detection success rate of 93%.

Once matured, this method can be implemented with current mapping coordinates systems and medical records softwares (e.g. Epic Systems). Over time, this machine-learning method is not solely used for ARPKD but may also detect early patient risks for various liver and bile duct cancers.