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

TL;DR

This work tackles the domain shift between public $3.0T$ prostate mp-MRI data and local $1.5T$ clinics by proposing a two-stage framework: first, unpaired image-to-image translation using an ACL-GAN to perform $3.0T$ to $1.5T$ domain transfer and augment training data; second, uncertainty-aware clinically significant PCa classification using Evidential Focal Loss with dataset filtering to suppress high-uncertainty samples. The approach yields improved downstream classification performance and provides predictive confidence estimates, addressing interpretability concerns in clinical deployment. Key contributions include a novel ACL-GAN–based domain-transfer pipeline, the Evidential Focal Loss that couples focal loss with Dirichlet-based uncertainty, and explicit data-filtering strategies (patch- and patient-driven) that improve robustness and calibration. The framework demonstrates substantial gains in AUC and calibration over baselines and holds promise for speeding up PCa diagnosis in data-constrained clinical settings by offering reliable uncertainty information to radiologists.

Abstract

Prostate Cancer (PCa) is a prevalent disease among men, and multi-parametric MRIs offer a non-invasive method for its detection. While MRI-based deep learning solutions have shown promise in supporting PCa diagnosis, acquiring sufficient training data, particularly in local clinics remains challenging. One potential solution is to take advantage of publicly available datasets to pre-train deep models and fine-tune them on the local data, but multi-source MRIs can pose challenges due to cross-domain distribution differences. These limitations hinder the adoption of explainable and reliable deep-learning solutions in local clinics for PCa diagnosis. In this work, we present a novel approach for unpaired image-to-image translation of prostate multi-parametric MRIs and an uncertainty-aware training approach for classifying clinically significant PCa, to be applied in data-constrained settings such as local and small clinics. Our approach involves a novel pipeline for translating unpaired 3.0T multi-parametric prostate MRIs to 1.5T, thereby augmenting the available training data. Additionally, we introduce an evidential deep learning approach to estimate model uncertainty and employ dataset filtering techniques during training. Furthermore, we propose a simple, yet efficient Evidential Focal Loss, combining focal loss with evidential uncertainty, to train our model effectively. Our experiments demonstrate that the proposed method significantly improves the Area Under ROC Curve (AUC) by over 20% compared to the previous work. Our code is available at https://github.com/med-i-lab/DT_UE_PCa

Figures (6)

• Figure 1: Visualization of sample data. \ref{['fig:t2kgh']} and \ref{['fig:adckgh']} are the 1.5T T2 and ADC images from KHSC, respectively. Similarly, \ref{['fig:t2px']} and \ref{['fig:adcpx']} are the 3.0T T2 and ADC images from the "ProstateX" Challenge, respectively.
• Figure 2: Detailed schematic of the proposed method. The overall framework of our proposed method contains two stages: 1), domain translation to map public 3.0T MRI with local 1.5T MRI; 2), uncertainty-aware clinically significant PCa classification. The bottom figure is the training schema for domain transfer. The upper right portion of the figure illustrates the PCa classification training process, which involves training the classifier using the Evidential Focal loss, filtering the training set based on uncertainty, and retraining the classifier on the filtered data to obtain the final classifier.
• Figure 3: Detailed architecture of "M.S. MpMRI" model. The first sequence of CNN layers contains 1 $\times$ 3D convolution layer and 4 $\times$ 2D convolution layers, 2 $\times$ Max Pooling layers with window size $2 \times 2$. Both extracted feature maps of T2 and ADC are concatenated channel-wise. After that, another set of convolution-max pooling layers is utilized. Finally, the extracted 2D features are reshaped to 1D and fed into a Fully connected layer follow by a softmax layer with 2 outputs representing the probabilities of which class the input data belongs to.
• Figure 4: Test performance of selected models based on uncertainty threshold.
• Figure 5: Both figures demonstrate the original AUC results. \ref{['fig:auc_old']} shows the comparison of AUC curves between the baseline and the models without filtration (experiments in the first category); \ref{['fig:auc_new']} shows the comparison of AUC curves between the baseline, the best model without filtration, and the best model with filtration on the training set. "EFL" is short for Evidential Focal Loss. The shaded areas in both figures represent the 95% confidence intervals (CI) of each model. CIs are obtained by using Bootstrap with $n = 3000$.