Abstract: While deep learning methods have shown great promise in improving the effectiveness of prostate cancer (PCa) diagnosis by detecting suspicious lesions from trans-rectal ultrasound (TRUS), they must overcome multiple simultaneous challenges. There is high heterogeneity in tissue appearance, significant class imbalance in favor of benign examples, and scarcity in the number and quality of ground truth annotations available to train models. Failure to address even a single one of these problems can result in unacceptable clinical outcomes.
Introduction
Prostate cancer (PCa) diagnosis is shifting toward targeted biopsies, where clinicians aim for specific suspicious lesions identified through imaging. While multi-parametric MRI (mp-MRI) is currently the gold standard, its high cost and requirement for specialized facilities limit its accessibility. Micro-ultrasound (micro-US), operating at a high resolution of 29 MHz, offers a real-time, cost-effective alternative. However, interpreting these high-resolution images is challenging even for experts.
Deep learning (DL) could bridge this gap, but it faces four major hurdles in this domain: weak labeling (biopsy results apply to tissue cores, not specific pixels), label scarcity, significant class imbalance (benign samples vastly outnumber cancer), and data heterogeneity across different clinical centers. The TRUSWorthy framework is designed to address these challenges simultaneously to provide a clinically viable detection system.
Methodology
The TRUSWorthy pipeline integrates several modern computer vision techniques tailored for medical imaging to handle the complexities of ultrasound data.
The framework consists of four primary components:
- Self-Supervised Learning (SSL): Using the VICReg framework, a ResNet-18 backbone is pre-trained on unlabeled micro-US data. This allows the model to learn the fundamental textures of ultrasound before being exposed to sparse cancer labels.
- Multiple Instance Learning (MIL) with Transformers: To address weak labeling, each biopsy core is treated as a "bag" of 55 small patches (ROIs). A Transformer aggregator processes these patches simultaneously to identify suspicious regions and classify the entire core.
- Mixed Deep Ensembles: Inspired by the RUSBoost strategy, each member of the ensemble is trained on all available cancer cases but a different random subset of benign cases. This diversity helps the ensemble better distinguish malignant tissue from varied benign presentations.
- Uncertainty Calibration: By averaging predictions across this diverse ensemble, the system calculates a confidence score. This enables a "reject" option, allowing the system to signal a human doctor when it is uncertain.
Datasets and Validation
The study utilized a substantial multi-center dataset comprising 693 patients and 6,607 biopsy cores from five distinct clinical centers.
Validation was performed using 5-fold cross-validation and a "leave-one-center-out" approach to ensure the model's ability to generalize to entirely new clinical environments. The dataset includes ground-truth histopathology (Gleason Scores) for all cores.
Key Results and Insights
TRUSWorthy achieved state-of-the-art performance with an AUROC of 79.9% and a Balanced Accuracy of 71.5%, outperforming specialized models like TRUSformer.
- The Power of the "Reject" Option: A critical finding was that as the system's rejection rate increased (discarding the most uncertain predictions), accuracy climbed dramatically. When allowed to reject the 80% of cases it was least sure about, accuracy on the remaining "top tier" predictions reached over 91%.
- Calibration and Trust: The system demonstrated significantly lower Expected Calibration Error (ECE), meaning its predicted probabilities closely align with real-world frequencies—a vital trait for building clinician trust.
- SSL as a Regularizer: SSL proved critical for the Transformer-based MIL to function effectively; without it, the model tended to overfit the limited, noisy labels of the training set.
Conclusion
TRUSWorthy represents a significant step toward the clinical application of AI in prostate cancer detection. By addressing data scarcity, imbalance, and uncertainty, it provides a reliable tool for clinicians. While the results are promising, future work will focus on expanding the dataset to include the transition zone and anterior gland, as well as conducting prospective clinical trials to validate the system's real-time efficacy during biopsies.
Summary updated