Classification of Cancer at Prostate MRI: Deep Learning versus Clinical PI-RADS Assessment

Abstract

R ecent highest-level evidence is mounting that there is an unequivocal benefit of MRI-targeted biopsies replacing or complementing systematic biopsies for diagnosis of prostate cancer (1-3). Realizing the integral role of MRI in diagnosis of prostate cancer, the Prostate Imaging Reporting and Data System (PI-RADS) continues to be developed (4-7). PI-RADS allows standardization of prostate MRI interpretation, which is a difficult task due to heterogeneous signal changes from benign prostatic hyperplasia, inflammation, and scarring after biopsy mimicking or hiding the appearance of prostate cancer. Even with PI-RADS, the high level of expertise required for accurate interpretation and persistent interobserver variability (8) limit consistency and availability while the demand for prostate MRI interpretation is at unprecedented levels. As convolutional neural networks approach or exceed human performance in natural image analysis (9), they promise to revolutionize computer-aided diagnosis and are increasingly evaluated in prostate MRI (10-15). U-Net is a convolutional neural network architecture optimized for image segmentation. It consists of an encoder and decoder

Comparison of Prostate MRI Lesion Segmentation Agreement Between Multiple Radiologists and a Fully Automatic Deep Learning System

“…In addition, probability maps were provided by U-Net. Briefly, U-Net was previously trained and validated for prostate and lesion segmentation using a cohort of 250 examinations for training and cross-validation and another cohort of 62 examinations for independent testing [11] and has demonstrated comparable performance to clinical PI-RADS interpretation. For 134 of the 165 included examinations that were part of the original training set, the four U-Nets of the full ensemble of 16 U-Nets that had not been trained with each respective case were used to calculate average probability maps for each patient.…”

“…For the remaining 31 examinations that were part of the original test set, all 16 U-Net members of the ensemble contributed to the probability maps. Probability maps were thresholded at 0.22, corresponding to the threshold mimicking a PI-RADS 3 assessment established during definition and validation of the U-Net [11]. In the resulting binary images, non-contiguous regions were separated, and each isolated region considered as a separate lesion segmentation.…”

“…This requires evaluation of clinical performance in consecutive patients screened for sPC also including patients that are ultimately not diagnosed with sPC [11], rather than evaluating pre-selected cohorts such as only patients with visible MR lesions [12]. Systems that utilize bi-parametric (T2-weighted and diffusion-weighted) MRI [11] are potentially able to extract more information than systems focusing only on a single sequence [12], as the multiparametric approach to prostate MRI has long been known to benefit from the combination of high-resolution anatomical and information-rich functional imaging [14]. It is generally expected that increasing amounts of data for the training of neural networks will improve the quality and generalizability of these models.…”

Detecting Prostate Cancer with Deep Learning for MRI: A Small Step Forward

“…A positive result of Schelb et al (10) was the high similarity of the AI algorithm to segment the entire gland comparable to manual segmentations. This opens the possibility for segmentation of the prostate gland to be used immediately for fusion biopsy and radiation therapy planning.…”

“…In this issue of Radiology, Schelb et al (10) demonstrate the potential of deep learning approaches to provide diagnostic support to radiologists for image interpretation and target delineation by using diagnostic biparametric MRI. The current work included a curated sample of 250 patients, representative of a mixed outpatient urologic clinic (biopsynaïve patients, patients with negative findings from prior biopsy, and patients treated with active surveillance).…”