A Deep Learning Approach to Diagnostic Classification of Prostate Cancer Using Pathology-Radiology Fusion

Abstract

Background: A definitive diagnosis of prostate cancer requires a biopsy to obtain tissue for pathologic analysis, but this is an invasive procedure and is associated with complications.

Purpose: To develop an artificial intelligence (AI)-based model (named AI-biopsy) for the early diagnosis of prostate cancer using magnetic resonance (MR) images labeled with histopathology information.

Study type: Retrospective.

Population: Magnetic resonance imaging (MRI) data sets from 400 patients with suspected prostate cancer and with histological data (228 acquired in-house and 172 from external publicly available databases).

Field strength/sequence: 1.5 to 3.0 Tesla, T2-weighted image pulse sequences.

Assessment: MR images reviewed and selected by two radiologists (with 6 and 17 years of experience). The patient images were labeled with prostate biopsy including Gleason Score (6 to 10) or Grade Group (1 to 5) and reviewed by one pathologist (with 15 years of experience). Deep learning models were developed to distinguish 1) benign from cancerous tumor and 2) high-risk tumor from low-risk tumor.

Statistical tests: To evaluate our models, we calculated negative predictive value, positive predictive value, specificity, sensitivity, and accuracy. We also calculated areas under the receiver operating characteristic (ROC) curves (AUCs) and Cohen's kappa.

Results: Our computational method (https://github.com/ih-lab/AI-biopsy) achieved AUCs of 0.89 (95% confidence interval [CI]: [0.86-0.92]) and 0.78 (95% CI: [0.74-0.82]) to classify cancer vs. benign and high- vs. low-risk of prostate disease, respectively.

Data conclusion: AI-biopsy provided a data-driven and reproducible way to assess cancer risk from MR images and a personalized strategy to potentially reduce the number of unnecessary biopsies. AI-biopsy highlighted the regions of MR images that contained the predictive features the algorithm used for diagnosis using the class activation map method. It is a fully automatic method with a drag-and-drop web interface (https://ai-biopsy.eipm-research.org) that allows radiologists to review AI-assessed MR images in real time.

Level of evidence: 1 TECHNICAL EFFICACY STAGE: 2.

Keywords: MRI images; PI-RADS; artificial intelligence; biopsy; deep neural networks; prostate cancer.

FIGURE 1

Method flow chart. (a) Unsegmented consistent sequences of seven axial T2w magnetic resonance (MR) image slices for each patient were selected that represent the prostate glands. (b) Each patient's MRI slice labeled by their corresponding biopsy result based on its Grade Group (GG) and Gleason Score (GS). (c) A convolutional neural network (CNN)‐based model (Model 1) classifies the cancer vs. benign and subsequently, and the second CNN‐based model (Model 2) predicts the risk level for each patient. (d) We highlighted the regions of MR images that algorithms focus on for prediction and compared the output of Model 2 with Prostate Imaging Reporting and Data System (PI‐RADS) using pathology labels as ground truth for a subset of test set. Receiver operating characteristic curves (ROCs) were used to assess the performance of different models based on individual patient.

FIGURE 2

Performance of two trained models for individual patient in the test set. (a) Model 1 performance for classifying cancer vs. benign. (b) The number of patients that were identified correct or incorrect by Model 1, negative predictive value, positive predictive value, specificity, sensitivity, and accuracy for cancer vs. benign. (c) Model 2 performance for classifying high risk vs. low risk. (d) The number of patients that were identified correct or incorrect by Model 2, negative predictive value, positive predictive value, specificity, sensitivity, and accuracy for high risk vs. low risk.

FIGURE 3

The highlighted prostate glands using class activation map (CAM) and radiologists. Model 2 classifies each image as high risk or low risk, and the deep feature analysis highlights the discriminative regions of the images. A radiologist marked the prostate gland of the images using green square dots. Biopsy results (based on Grade Groups [GGs]) as ground truth and Prostate Imaging Reporting and Data System (PI‐RADS) also are indicated in the figure. (a) Artificial intelligence (AI)‐biopsy predicts the risk level of cases (with a probability score for each class) and highlighted the prostate gland correctly. (b) AI‐biopsy is not able to predict the correct risk level of cases in which the prostate glands are not correctly detected. Red color illustrates features with higher weight.

FIGURE 4

AI‐biopsy is a fully automated framework to use in clinics for evaluation of the prostate cancer risk level. We employed a threshold condition on the output of both models for diagnosis using minimum seven T2w axial image slices. (a) While for prediction of benign diagnosis, all seven image slices should get P ≥ 0.5 for the benign class; (b) one image slice (out of seven imported image slices) with P ≥ 0.5 is enough for Model 1 to result in cancer prediction; (c) Model 2 needs at least two image slices (out of seven imported image slices) with high‐risk P ≥ 0.5 for a patient to result in high‐risk diagnosis; and (d) the result explanation could be seen by clicking on “N/A” option in the web interface (https://ai‐biopsy.eipm‐research.org).