Autosomal dominant polycystic kidney disease (ADPKD) is a genetic disorder characterised by the progressive formation of renal cysts, which ultimately impair kidney function. Traditionally, imaging techniques such as magnetic resonance imaging (MRI) have been employed to evaluate cystic burden. Still, manual segmentation is time-consuming and can vary in accuracy depending on the radiologist. A recent study reviewed in European Radiology Experimental introduces a deep learning-based approach to automate the segmentation of renal cysts on T2-weighted MRI scans in ADPKD patients, offering the potential for increased accuracy and efficiency in diagnostic and prognostic assessments.
The Growing Need for Automated Renal Imaging
Current diagnostic methods for ADPKD rely on manual segmentation to assess biomarkers like total kidney volume (TKV) and total cystic volume (TCV), both of which are crucial for predicting disease progression. High TKV, for example, has been linked to complications such as hypertension and haematuria, as well as an increased risk of kidney failure. The manual segmentation process, however, demands extensive expertise and time, with each kidney taking approximately one hour to segment. Given the increasing prevalence of ADPKD, there is a strong case for developing automated imaging solutions to optimise these evaluations.
Deep learning has emerged as a promising tool in medical imaging, with the potential to deliver automated, consistent results that align closely with human interpretation. By training on thousands of images, deep learning algorithms can identify and segment features like cysts and kidney outlines, bypassing much of the manual work traditionally needed. Nevertheless, for deep learning models to be truly effective in clinical settings, they must be trained with diverse datasets encompassing various ADPKD stages and different imaging protocols to reflect real-life practice conditions.
Training and Testing the Deep Learning Model
The study employed a deep learning model to segment renal cysts using a dataset of MRI images from ADPKD patients scanned between 2008 and 2022. Researchers used the most recent 20 MRI scans as a test dataset to simulate a real-world diagnostic setting, while the remaining scans formed the training dataset. Additionally, eight scans from individuals without ADPKD were included in the training data to aid in refining segmentation accuracy across various disease stages, including cases with lower cystic burden.
The model’s performance was compared against three human raters, each with varying levels of expertise in renal imaging, who manually segmented the same test dataset. This multi-rater approach was chosen to evaluate not only the algorithm’s accuracy but also its consistency within the range of human interpretation. To assess overlap and accuracy, researchers applied metrics such as the Dice similarity coefficient, which measures the similarity between the algorithm’s output and the ground-truth segmentations. Bland-Altman analysis was also used to examine differences in TKV and TCV estimations between the algorithm and the raters, highlighting any potential biases.
Overall, the algorithm’s performance fell within the range of inter-rater variability. The Dice similarity scores ranged from 85.9% to 87.4% when comparing the algorithm with human raters, indicating a level of agreement consistent with human performance. Although the algorithm performed well across most cases, it struggled with two patients whose TCV exceeded 2800 mL, where it tended to underestimate both TKV and TCV. This underestimation in cases of high cystic volume suggests a need for broader training data to account for advanced disease stages.
Challenges and Future Directions in Deep Learning for ADPKD
Despite promising results, this deep learning model faced challenges that are instructive for the future development of automated imaging solutions. One notable limitation was its performance in severe cases, where large cysts were under-represented in the training dataset. Consequently, the algorithm occasionally underestimated cystic volume in these cases, which could impact its reliability in advanced ADPKD patients. Including a more comprehensive dataset with a range of cyst sizes and higher TCV cases could help mitigate this issue, making the algorithm more adaptable to different stages of the disease.
Another challenge arose from the heterogeneous imaging protocols in the dataset, which included scans from various MRI vendors and field strengths. These differences can affect the model’s performance, as imaging parameters like magnetic field strength have been shown to influence deep learning algorithms’ accuracy in segmenting organs. Although the algorithm’s performance was relatively stable across different imaging conditions, it showed slightly lower accuracy in scans acquired at 3 Tesla (3T) compared to 1.5 Tesla (1.5T). Future work could involve optimising the algorithm for different field strengths and imaging settings, making it more versatile for use across different institutions.
Lastly, the segmentation of clustered or irregularly shaped cysts remains a complex task for the algorithm. The study revealed that the cystic index—a measure of cystic volume relative to TKV—was a more consistent metric than TCV alone, reducing variability both between the algorithm and human raters and among raters themselves. The algorithm’s use of the cystic index, therefore, could hold particular value for clinical application, as it offers a robust marker that may improve consistency in disease assessment.
The integration of deep learning into renal imaging for ADPKD patients represents a promising step towards enhancing diagnostic and prognostic precision. The studied algorithm demonstrates a strong capability to match human-level segmentation accuracy within the range of inter-rater variability, although some limitations remain, particularly in cases with large cystic volumes. Addressing these gaps by expanding training datasets and optimising performance across imaging conditions will be essential for real-world applications. With further refinement, automated segmentation could play a key role in enabling routine, large-scale monitoring of TKV and cystic index, ultimately supporting more timely and personalised treatment strategies for ADPKD patients.
Source: European Radiology Experimental
Image Credit: iStock
