Decoupled conditional contrastive learning with variable metadata for prostate lesion detection

Camille Ruppli, Pietro Gori, Roberto Ardon, Isabelle Bloch

Abstract

Detecting prostate lesions in multi-parametric MRI (mp-MRI) is a critical clinical task often hindered by the scarcity of high-quality, biopsy-confirmed labels and the high inter-reader variability of radiological scores. This paper presents a novel contrastive learning framework designed to leverage "noisy" clinical metadata by incorporating a confidence-weighted loss function. By decoupling the contrastive loss and weighting sample interactions based on radiologist agreement, the proposed method achieves significant performance gains, particularly in low-data regimes, outperforming standard self-supervised and supervised baselines.

Introduction

The standard of care for prostate cancer screening involves multi-parametric MRI (mp-MRI), which radiologists interpret using the Prostate Imaging-Reporting and Data System (PI-RADS). However, training deep learning models for automated detection faces two primary hurdles. First, the gold standard—pathological results from biopsies—is expensive to obtain and often biased toward high-risk patients. Second, while PI-RADS scores are more abundant, they are subject to significant inter-reader variability; different radiologists frequently assign different scores to the same lesion.

Traditional supervised learning models struggle to converge effectively when faced with such label noise, while standard self-supervised learning (SSL) methods typically ignore clinical metadata entirely. This research introduces a Decoupled Conditional Contrastive Learning approach that treats metadata not as a static label, but as a conditional signal weighted by the level of expert agreement.

Methodology

The core of the proposed framework is a hybrid loss function that combines unsupervised contrastive learning with a metadata-conditioned component. The technical approach consists of four primary pillars:

1. Metadata Binarization and Filtering

To standardize disparate scoring systems, the authors binarized PI-RADS and biopsy results into "healthy" (PI-RADS 1-2) and "malignant" (PI-RADS 4-5) categories. Notably, PI-RADS 3 cases—clinically ambiguous "gray area" cases—were excluded during the pre-training phase to prevent the model from learning from highly inconsistent signals.

2. Confidence Scoring via Agreement

For exams reviewed by multiple radiologists, the authors calculated a confidence score ($c$) based on majority voting. If all annotators agreed on a label, the confidence was set to 1.0. If there was a split in opinion, the confidence score was reduced proportionally. This allows the model to prioritize learning from "clear-cut" cases while remaining cautious with disputed ones.

3. The Weighted Kernel and Decoupling

In standard contrastive learning (e.g., SimCLR), every distinct image is treated as a negative pair and pushed apart in the latent space. This is problematic in medical imaging, where two different patients may have clinically identical tumors. The authors addressed this through:

  • Alignment: If two scans share the same high-confidence label, the model is encouraged to bring their representations together.
  • Decoupling: The loss function is "decoupled" to ensure that samples with identical metadata are not mistakenly repelled from one another, preserving the clinical similarity of the latent space.

4. Hybrid Loss Formulation

The final loss function integrates this conditional logic for annotated data with standard unsupervised contrastive loss for unannotated data. This allows the model to benefit from the entire available dataset, regardless of whether a specific scan has associated metadata.

Experimental Setup

The study utilized a robust dataset comprising:

  • Private Dataset: 2,415 mp-MRI exams, including 1,397 with manual segmentations and metadata provided by up to seven different radiologists.
  • PI-CAI Challenge Dataset: 1,500 public exams with 1,295 annotations.

The model input consisted of four registered volumes: T2-weighted (T2W), Apparent Diffusion Coefficient (ADC), high b-value Diffusion-Weighted Imaging (DWI), and a prostate zone segmentation map.

Key Results and Discussion

The framework was evaluated using 5-fold cross-validation, with a focus on the Area Under the Curve (AUC) for both exam-level and lesion-level detection.

Performance in Low-Data Regimes

The method demonstrated its greatest utility when labeled data was scarce. Using only 1% of the annotated data, the proposed framework achieved a 0.70 AUC on the PI-CAI dataset. This significantly outperformed random initialization (0.68) and established SSL methods like MoCo and BYOL. This suggests that incorporating "weak" metadata during pre-training can effectively bridge the gap created by a lack of expensive, pixel-perfect manual segmentations.

Ablation Studies

Ablation experiments confirmed that both the confidence weighting and the decoupling mechanism were essential. Removing the confidence scores led to a drop in performance, as the model was forced to treat inconsistent labels as absolute truths. Similarly, removing the decoupling caused the model to struggle with intra-class variance, as it tried to push similar clinical cases apart in the latent space.

Conclusion and Future Work

This research demonstrates that annotator disagreement should not be viewed merely as noise to be filtered, but as a valuable signal that reflects the inherent uncertainty of clinical diagnosis. By modeling this uncertainty through a decoupled conditional contrastive loss, the authors improved the robustness of prostate cancer detection models.

Future work will focus on integrating PI-RADS 3 cases into the training pipeline rather than excluding them, potentially using soft-labeling techniques. Additionally, the authors aim to address the computational challenges of 3D contrastive learning to better capture the spatial relationships within the prostate volume while maintaining efficiency.

Summary updated

BibTeX
@inproceedings{ruppli2023decoupled,
  author = {Camille Ruppli and Pietro Gori and Roberto Ardon and Isabelle Bloch},
  title = {{Decoupled conditional contrastive learning with variable metadata for prostate lesion detection}},
  booktitle = {MILLanD @ MICCAI},
  year = {2023},
  doi = {10.1007/978-3-031-44917-8\_9},
}
Figure