Deep learning model rivals radiologists in detecting prostate cancer with MRI

Recent Radiology This journal study evaluates the ability of a fully automated deep learning (DL) model to generate deterministic outputs for identifying clinically significant prostate cancer (csPCa).

study: A fully automated deep learning model for detecting clinically significant prostate cancer on MRI. Image credit: Antonio Marca / Shutterstock.com

Diagnosing Prostate Cancer Using Machine Learning

Prostate cancer is the second most common cancer occurring in men worldwide, and multiparametric magnetic resonance imaging (MRI) is commonly used to diagnose csPCa.

Standardized reporting and interpretation approaches include the use of the Prostate Imaging Reporting and Data System (PI-RADS), which requires a high level of expertise, but using PI-RADS to classify lesions is susceptible to intraobserver and interobserver variability.

Traditional machine learning or DL ​​can be used to detect csPCa by training a model on specific regions of interest informed by MRI scans. Another approach is to train a segmentation model to obtain predictions for each voxel.

These machine learning approaches require radiologists or pathologists to annotate lesions, not only during the model development phase, but also during retraining and reevaluation phases after clinical implementation, which results in high implementation costs for these approaches and limits the size of the datasets.

About the Research

The researchers in this study were interested in developing a DL model to predict the presence of csPCa without prior information about the location of the tumor. They used patient-level labels that revealed the presence or absence of csPCa and compared the model's predictions to those of radiologists.

Data were collected from patients without known csPCa who underwent MRI scans between January 2017 and December 2019. T1-weighted contrast-enhanced images, T2-weighted images, apparent diffusion coefficient maps, and diffusion-weighted images were used to train a convolutional neural network to predict csPCa.

Pathological diagnosis formed the standard.Four models were evaluated: image only, radiologist, image + radiologist, and image + clinical + radiologist.

The PI-RADS assessment of the four radiologists was reflected in the external (ProstateX) test set and was also used for the internal test set. DeLong test and receiver operating characteristic curve (AUC) were used to evaluate the radiologists' performance. Tumor location was delineated using gradient-weighted class activation maps (Grad-CAM).

Research findings

The image + clinical + radiologist model was associated with the highest predictive power with an AUC of 0.94, followed by the image + clinical model with an AUC of 0.91. The image only model vs. radiologist had an AUC of 0.89.

In the subset of pathologically proven cases in the internal set, the image + clinical model had the highest AUC of 0.88. The radiologist model had an AUC of 0.78, while the clinical benchmark had an AUC of 0.77. Thus, the image + clinical + radiologist model had the highest predictive power across the entire internal test sample. In contrast, in the subset of pathologically proven cases, the image + clinical model had the highest predictive power.

The image + clinical + radiologist model had the highest true positive rate (TPR) and the lowest false positive rate (FPR). For pathologically proven cases, radiologist had the highest TPR and the image + clinical model had the lowest FPR. For the external dataset, the image + radiologist model had the highest AUC and TPR and the lowest FPR.

Regarding the use of Grad-CAM for tumor localization, patients with PI-RADS 1 or 2 lesions and biopsy This accounted for a significant proportion of negative cases, with some cases classified as false negatives.

Conclusion

In this study, we successfully predicted the presence of csPCa on MRI using a DL model. No statistically significant difference was observed between the performance of our model and that of experienced radiologists in both the internal and external test sets. These results indicate that the DL model developed in this study may assist radiologists in identifying csPCa and lesion biopsy, significantly improving the diagnosis of prostate cancer.

An important limitation of this study is that it is a single-center, retrospective study. Furthermore, to improve prediction accuracy, the DL model only included radiologists who specialize in prostate MRI and excluded residents and general radiologists.