Domain Transfer Through Image-to-Image Translation for Uncertainty-Aware Prostate Cancer Classification

Meng Zhou, Amoon Jamzad, Jason Izard, Alexandre Menard, Robert Siemens, Parvin MousaviarXiv 2023 · DOI · arXiv

Introduction

Deep learning has demonstrated significant potential in automating the detection of Prostate Cancer (PCa) using multi-parametric MRI (mp-MRI). However, clinical adoption is often hindered by two critical challenges: the domain gap between different imaging hardware and the lack of interpretability regarding model confidence. This research introduces a two-stage framework designed to bridge the gap between high-field (3.0T) and low-field (1.5T) MRI domains while providing reliable, uncertainty-aware classifications. By combining GAN-based image translation with Evidential Deep Learning, the proposed method achieves state-of-the-art performance on local clinical data.

The Problem: Domain Mismatch and the "Black Box"

In the real world, medical imaging models face significant deployment hurdles:

  1. The Domain Gap: Large, high-quality public datasets are typically acquired using high-end 3.0T MRI scanners. In contrast, many smaller or local clinics utilize 1.5T scanners for cost-effectiveness. Models trained on 3.0T data often fail to generalize to 1.5T data due to variations in noise, resolution, and contrast.
  2. The "Black Box" Problem: Standard deep learning models provide predictions without a measure of confidence. In a clinical setting, a "wrong but confident" prediction can lead to misdiagnosis. Radiologists require a system that can flag "unsure" cases for manual review.

Methodology

The authors propose a two-stage pipeline to address these issues through domain adaptation and evidential reasoning.

Stage 1: Domain Transfer via ACL-GAN

To mitigate the domain gap, the framework employs the Adversarial Consistency Loss GAN (ACL-GAN) for unpaired image-to-image translation. This model is trained to "translate" 3.0T images from public datasets into the 1.5T domain of the local clinic. Unlike traditional CycleGANs, which can be overly restrictive, ACL-GAN utilizes a distribution-based consistency loss. This ensures that essential anatomical structures are preserved while the visual characteristics (noise and contrast) are successfully mapped to the target 1.5T domain.

Stage 2: Uncertainty-Aware Classification

Once the data is translated, it is processed by a classification network designed for reliability:

  • Evidential Deep Learning (EDL): Instead of standard softmax probabilities, the model outputs parameters of a Dirichlet distribution. This allows the system to quantify "vacuous uncertainty"—a measure of whether the model has seen enough similar data to make a reliable prediction.
  • Evidential Focal Loss (EFL): The authors introduced a novel loss function that integrates Focal Loss (to handle class imbalance) with the evidential framework, focusing the model on "hard" samples during training.
  • Uncertainty-Based Filtering: The model identifies training samples with high uncertainty, which often represent noisy labels or outliers. By filtering out the most uncertain 20% of image patches and retraining on the "clean" data, the model's robustness is significantly enhanced.

Experimental Setup and Datasets

The study utilized two primary data sources:

  • ProstateX Grand Challenge (Public): A 3.0T MRI dataset (T2-weighted and ADC maps) from 204 patients, serving as the source domain.
  • Kingston Health Science Center (KHSC) (Local): A private 1.5T MRI dataset from 104 patients. 70 patients were used for training/validation, and 34 were reserved as a standalone test set to simulate real-world clinical deployment.

Key Results

The proposed framework demonstrated a substantial performance leap over existing methods like SDNet:

  • Diagnostic Accuracy: The method achieved an Area Under the ROC Curve (AUC) of 98.4%, a massive improvement over the baseline's 76.2%.
  • Sensitivity: In several test scenarios, the model reached 100% sensitivity, ensuring no clinically significant prostate cancer cases were missed.
  • Calibration: The model exhibited low Expected Calibration Error (ECE), meaning its predicted confidence levels were highly representative of its actual accuracy.
  • Impact of Filtering: The uncertainty-based patch filtering was pivotal, raising the AUC from approximately 89% to over 98%.

Discussion and Future Work

The results suggest that explicit domain translation is superior to simple fine-tuning for cross-hardware applications. Furthermore, using uncertainty as a quality control tool acts as an automated data cleaning process, leading to sharper classification boundaries. This system supports a "human-in-the-loop" workflow, where the model flags ambiguous cases for expert radiologist review.

Despite these successes, the current framework is limited to 2D slice processing, which ignores the 3D spatial context of the prostate volume. Future research will focus on moving toward 3D volumetric translation and integrating additional functional sequences (such as $K^{trans}$) to further improve diagnostic precision.

Summary updated

BibTeX
@misc{zhou2023domain,
  author = {Meng Zhou and Amoon Jamzad and Jason Izard and Alexandre Menard and Robert Siemens and Parvin Mousavi},
  title = {{Domain Transfer Through Image-to-Image Translation for Uncertainty-Aware Prostate Cancer Classification}},
  year = {2023},
  eprint = {2307.00479},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.00479},
}
Figure