Bridging the gap between prostate radiology and pathology through machine learning.

BACKGROUND: Prostate cancer remains the second deadliest cancer for American men despite clinical advancements. Currently, magnetic resonance imaging (MRI) is considered the most sensitive non-invasive imaging modality that enables visualization, detection, and localization of prostate cancer, and is increasingly used to guide targeted biopsies for prostate cancer diagnosis. However, its utility remains limited due to high rates of false positives and false negatives as well as low inter-reader agreements. PURPOSE: Machine learning methods to detect and localize cancer on prostate MRI can help standardize radiologist interpretations. However, existing machine learning methods vary not only in model architecture, but also in the ground truth labeling strategies used for model training. We compare different labeling strategies and the effects they have on the performance of different machine learning models for prostate cancer detection on MRI. METHODS: Four different deep learning models (SPCNet, U-Net, branched U-Net, and DeepLabv3+) were trained to detect prostate cancer on MRI using 75 patients with radical prostatectomy, and evaluated using 40 patients with radical prostatectomy and 275 patients with targeted biopsy. Each deep learning model was trained with four different label types: pathology-confirmed radiologist labels, pathologist labels on whole-mount histopathology images, and lesion-level and pixel-level digital pathologist labels (previously validated deep learning algorithm on histopathology images to predict pixel-level Gleason patterns) on whole-mount histopathology images. The pathologist and digital pathologist labels (collectively referred to as pathology labels) were mapped onto pre-operative MRI using an automated MRI-histopathology registration platform. RESULTS: Radiologist labels missed cancers (ROC-AUC: 0.75-0.84), had lower lesion volumes (~68% of pathology lesions), and lower Dice overlaps (0.24-0.28) when compared with pathology labels. Consequently, machine learning models trained with radiologist labels also showed inferior performance compared to models trained with pathology labels. Digital pathologist labels showed high concordance with pathologist labels of cancer (lesion ROC-AUC: 0.97-1, lesion Dice: 0.75-0.93). Machine learning models trained with digital pathologist labels had the highest lesion detection rates in the radical prostatectomy cohort (aggressive lesion ROC-AUC: 0.91-0.94), and had generalizable and comparable performance to pathologist label-trained-models in the targeted biopsy cohort (aggressive lesion ROC-AUC: 0.87-0.88), irrespective of the deep learning architecture. Moreover, machine learning models trained with pixel-level digital pathologist labels were able to selectively identify aggressive and indolent cancer components in mixed lesions on MRI, which is not possible with any human-annotated label type. CONCLUSIONS: Machine learning models for prostate MRI interpretation that are trained with digital pathologist labels showed higher or comparable performance with pathologist label-trained models in both radical prostatectomy and targeted biopsy cohort. Digital pathologist labels can reduce challenges associated with human annotations, including labor, time, inter- and intra-reader variability, and can help bridge the gap between prostate radiology and pathology by enabling the training of reliable machine learning models to detect and localize prostate cancer on MRI.

The stanford prostate cancer calculator: Development and external validation of online nomograms incorporating PIRADS scores to predict clinically significant prostate cancer.

BACKGROUND: While multiparametric MRI (mpMRI) has high sensitivity for detection of clinically significant prostate cancer (CSC), false positives and negatives remain common. Calculators that combine mpMRI with clinical variables can improve cancer risk assessment, while providing more accurate predictions for individual patients. We sought to create and externally validate nomograms incorporating Prostate Imaging Reporting and Data System (PIRADS) scores and clinical data to predict the presence of CSC in men of all biopsy backgrounds. METHODS: Data from 2125 men undergoing mpMRI and MR fusion biopsy from 2014 to 2018 at Stanford, Yale, and UAB were prospectively collected. Clinical data included age, race, PSA, biopsy status, PIRADS scores, and prostate volume. A nomogram predicting detection of CSC on targeted or systematic biopsy was created. RESULTS: Biopsy history, Prostate Specific Antigen (PSA) density, PIRADS score of 4 or 5, Caucasian race, and age were significant independent predictors. Our nomogram-the Stanford Prostate Cancer Calculator (SPCC)-combined these factors in a logistic regression to provide stronger predictive accuracy than PSA density or PIRADS alone. Validation of the SPCC using data from Yale and UAB yielded robust AUC values. CONCLUSIONS: The SPCC combines pre-biopsy mpMRI with clinical data to more accurately predict the probability of CSC in men of all biopsy backgrounds. The SPCC demonstrates strong external generalizability with successful validation in two separate institutions. The calculator is available as a free web-based tool that can direct real-time clinical decision-making.

Deep Learning Improves Speed and Accuracy of Prostate Gland Segmentations on Magnetic Resonance Imaging for Targeted Biopsy.

PURPOSE: Targeted biopsy improves prostate cancer diagnosis. Accurate prostate segmentation on magnetic resonance imaging (MRI) is critical for accurate biopsy. Manual gland segmentation is tedious and time-consuming. We sought to develop a deep learning model to rapidly and accurately segment the prostate on MRI and to implement it as part of routine magnetic resonance-ultrasound fusion biopsy in the clinic. MATERIALS AND METHODS: A total of 905 subjects underwent multiparametric MRI at 29 institutions, followed by magnetic resonance-ultrasound fusion biopsy at 1 institution. A urologic oncology expert segmented the prostate on axial T2-weighted MRI scans. We trained a deep learning model, ProGNet, on 805 cases. We retrospectively tested ProGNet on 100 independent internal and 56 external cases. We prospectively implemented ProGNet as part of the fusion biopsy procedure for 11 patients. We compared ProGNet performance to 2 deep learning networks (U-Net and holistically-nested edge detector) and radiology technicians. The Dice similarity coefficient (DSC) was used to measure overlap with expert segmentations. DSCs were compared using paired t-tests. RESULTS: ProGNet (DSC=0.92) outperformed U-Net (DSC=0.85, p <0.0001), holistically-nested edge detector (DSC=0.80, p <0.0001), and radiology technicians (DSC=0.89, p <0.0001) in the retrospective internal test set. In the prospective cohort, ProGNet (DSC=0.93) outperformed radiology technicians (DSC=0.90, p <0.0001). ProGNet took just 35 seconds per case (vs 10 minutes for radiology technicians) to yield a clinically utilizable segmentation file. CONCLUSIONS: This is the first study to employ a deep learning model for prostate gland segmentation for targeted biopsy in routine urological clinical practice, while reporting results and releasing the code online. Prospective and retrospective evaluations revealed increased speed and accuracy.

Prostate Magnetic Resonance Imaging Interpretation Varies Substantially Across Radiologists.

BACKGROUND: Multiparametric magnetic resonance imaging (mpMRI) interpreted by experts is a powerful tool for diagnosing prostate cancer. However, the generalizability of published results across radiologists of varying expertise has not been verified. OBJECTIVE: To assess variability in mpMRI reporting and diagnostic accuracy across radiologists of varying experience in routine clinical care. DESIGN, SETTING, AND PARTICIPANTS: Men who underwent mpMRI and MR-fusion biopsy between 2014-2016. Each MRI scan was read by one of nine radiologists using the Prostate Imaging Reporting and Data System (PIRADS) and was not re-read before biopsy. Biopsy histopathology was the reference standard. OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS: Outcomes were the PIRADS score distribution and diagnostic accuracy across nine radiologists. We evaluated the association between age, prostate-specific antigen, PIRADS score, and radiologist in predicting clinically significant cancer (Gleason ≥7) using multivariable logistic regression. We conducted sensitivity analyses for case volume and changes in accuracy over time. RESULTS AND LIMITATIONS: We analyzed data for 409 subjects with 503 MRI lesions. While the number of lesions (mean 1.2 lesions/patient) did not differ across radiologists, substantial variation existed in PIRADS distribution and cancer yield. The significant cancer detection rate was 3-27% for PIRADS 3 lesions, 23-65% for PIRADS 4, and 40-80% for PIRADS 5 across radiologists. Some 13-60% of men with a PIRADS score of <3 on MRI harbored clinically significant cancer. The area under the receiver operating characteristic curve varied from 0.69 to 0.81 for detection of clinically significant cancer. PIRADS score (p<0.0001) and radiologist (p=0.042) were independently associated with cancer in multivariable analysis. Neither individual radiologist volume nor study period impacted the results. MRI scans were not retrospectively re-read by all radiologists, precluding measurement of inter-observer agreement. CONCLUSIONS: We observed considerable variability in PIRADS score assignment and significant cancer yield across radiologists. We advise internal evaluation of mpMRI accuracy before widespread adoption. PATIENT SUMMARY: We evaluated the interpretation of multiparametric magnetic resonance imaging of the prostate in routine clinical care. Diagnostic accuracy depends on the Prostate Imaging Reporting and Data System score and the radiologist.

Imaging of traumatic adrenal injury.

Traumatic adrenal injury is a relatively rare event, reported in 0.15 to 4 % of blunt abdominal trauma cases. The location of the adrenal glands, protected deeply within the retroperitoneum, accounts for the rarity of adrenal trauma. These injuries are unilateral in 75 to 90 % of cases and most commonly afflict the right adrenal gland. While no specific clinical symptoms or signs point directly to adrenal injury, and isolated adrenal injury is rare, the presence of adrenal injury can be an indicator of severe trauma. In fact, mortality rates in series of adrenal injuries range from 7 to 32 %. The most common associated injuries affect the liver, ribs, kidney, or spleen. Three theories of adrenal injury from blunt trauma have been proposed: (1) acute rise in intra-adrenal venous pressure due to compression of the IVC during impact, (2) crushing between the spine and surrounding organs, and (3) deceleration forces shearing the small adrenal arterioles. The most common imaging features include a 2-3-cm oval hematoma, irregular hemorrhage obliterating the adrenal gland, periadrenal hemorrhage or fat stranding, and uniform adrenal swelling with increased attenuation. The differential diagnosis of adrenal abnormalities on trauma CT includes adenoma, carcinoma, myelolipoma, metastases, pheochromocytoma, and tuberculosis. Preexisting adrenal disorders can predispose the adrenal to injury with minor trauma. Most adrenal traumatic injuries are managed conservatively.

Esophageal varices in cirrhotic patients: evaluation with liver CT.

OBJECTIVE: The objective of this study was to evaluate the performance of routine helical liver CT in the detection and grading of esophageal varices in cirrhotic patients. MATERIALS AND METHODS: A total of 67 consecutive cirrhotic patients who underwent both upper endoscopy and helical liver CT within a 4-week interval were evaluated. The CT protocol included unenhanced, arterial, and portal phases with a collimation of 7-7.5 mm. Two blinded abdominal imagers (6 and 7 years' experience) retrospectively interpreted all CT images to detect the presence of esophageal varices on a 5-point confidence scale and measure the largest varix identified. Receiver operating characteristic (ROC) curve analysis was performed, and the correlation between CT measurements and endoscopic grading, the reference standard, was assessed. RESULTS: The variceal detection rates for the observers was 92% (11/12) and 92% (11/12) for large (i.e., clinically significant) varices, 53% (16/30) and 60% (18/30) for small varices, and 64% (27/42) and 69% (29/42) for all varices. The area under the ROC curve for the detection of esophageal varices of any size was 0.77 (observer 1) and 0.80 (observer 2). CT variceal grading showed a strong correlation with endoscopic grading for both observers (p < or = 0.001). Using a variceal diameter threshold of 3 mm on CT, sensitivity, specificity, and accuracy for distinguishing large esophageal varices from small or no varices were 92% (11/12), 84% (46/55), and 85% (57/67), respectively, for both observers. CONCLUSION: Liver CT is useful for the detection and grading of esophageal varices. A diameter of 3 mm may be an appropriate screening threshold for large clinically significant varices.

Pancreatic and peripancreatic diseases mimicking primary pancreatic neoplasia.

A variety of anatomic variants and pathologic conditions in and around the pancreas may simulate primary pancreatic neoplasia at routine abdominal cross-sectional imaging. An ambiguous lesion whose appearance suggests a pancreatic origin requires a broad differential diagnosis that can subsequently be narrowed on the basis of both clinical history and features at optimal computed tomography (CT) and magnetic resonance (MR) imaging. Pancreas-specific multidetector CT and MR imaging techniques with thin collimation, multiplanar and multiphasic scans, and newly introduced curved planar reformation may help avoid potential diagnostic pitfalls. These techniques can help identify and characterize a mass in multiple viewing planes, thereby helping distinguish a true pancreatic neoplasm from peripancreatic adenopathy or from a tumor of the adjacent duodenum or small bowel. They can also help determine the cause of a tumor. It is important that the radiologist be familiar with the wide spectrum of anatomic variants and disease entities that can mimic primary pancreatic neoplasia in order to initiate the appropriate lesion-specific work-up and treatment and avoid unnecessary tests or procedures, including surgery.