Poisson Ordinal Network for Gleason Group Estimation Using Bi-Parametric MRI

TL;DR

The work addresses estimating Gleason groups from multi-parametric MRI (Gleason groups $1$–$5$) to reduce biopsy load in prostate cancer. It introduces the Poisson Ordinal Network (PON), modeling predictions with a Poisson distribution parameterized by $\lambda(\mathbf{x})$, coupling Poisson encoding of labels with Poisson focal loss to capture ordinal dependencies. A memory-bank-based contrastive learning module regularizes intra-/inter-class variance, enabling robust representations even with MR heterogeneity. On the PROMIS mpMRI dataset of 265 cases, PON achieves state-of-the-art results for Gleason-group estimation and clinically significant cancer detection, supporting potential clinical utility as a noninvasive triage aid.

Abstract

The Gleason groups serve as the primary histological grading system for prostate cancer, providing crucial insights into the cancer's potential for growth and metastasis. In clinical practice, pathologists determine the Gleason groups based on specimens obtained from ultrasound-guided biopsies. In this study, we investigate the feasibility of directly estimating the Gleason groups from MRI scans to reduce otherwise required biopsies. We identify two characteristics of this task, ordinality and the resulting dependent yet unknown variances between Gleason groups. In addition to the inter- / intra- observer variability in a multi-step Gleason scoring process based on the interpretation of Gleason patterns, our MR-based prediction is also subject to specimen sampling variance and, to a lesser degree, varying MR imaging protocols. To address this challenge, we propose a novel Poisson ordinal network (PON). PONs model the prediction using a Poisson distribution and leverages Poisson encoding and Poisson focal loss to capture a learnable dependency between ordinal classes (here, Gleason groups), rather than relying solely on the numerical ground-truth (e.g. Gleason Groups 1-5 or Gleason Scores 6-10). To improve this modelling efficacy, PONs also employ contrastive learning with a memory bank to regularise intra-class variance, decoupling the memory requirement of contrast learning from the batch size. Experimental results based on the images labelled by saturation biopsies from 265 prior-biopsy-blind patients, across two tasks demonstrate the superiority and effectiveness of our proposed method.

Figures (3)

• Figure 1: (a) Pathologists assign Gleason groups based on how the cancer cells look like healthy tissue under the microscope. Although cell growth is a continuous process, pathologists discretize it (dotted lines) into multiple groups. (b) Both predictions are incorrect and yield the same cross-entropy loss. However, a multi-modal distribution tends to produce misclassifications that deviate further from the label, resulting in different treatments. Consequently, it poses a higher clinical risk. Consider an image predicted as 3+3. The multi-modal distribution appears counter-intuitive, as pathologists will not assign second highest confidence to >4+3 due to the obvious difference.
• Figure 2: (a) Our method represents the prediction probability using a Poisson distribution with the parameter $\lambda(\textbf{x})$ output from the classifier, and stores the output from projector in a memory bank. (b) For ordinal classification, we introduce the inherent class ordering through Poisson encoding. (c) To regulate the inter-/intra-class covariance, we conduct contrastive learning between the sample and the memory bank, which pulls together features with the same label and pushes away different labeled features
• Figure 3: Left: our method helps distinguish 30 negatives in 48 false positive samples under the same sensitivity, and 31 positive samples in 52 false negative samples under the same specificity by radiologists. Right: Two examples of the network predicting correctly and radiologists predicting wrong. Arrows indicate lesions.