Abstract
Deep learning (DL) is a rapidly developing field in machine learning (ML). The concept of deep learning originates from research on artificial neural networks and is an upgrade of traditional neural networks. It has achieved great success in various domains and has shown potential in solving medical problems, particularly when using medical images. Bladder cancer (BCa) is the tenth most common cancer in the world. Imaging, as a safe, noninvasive, and relatively inexpensive technique, is a powerful tool to aid in the diagnosis and treatment of bladder cancer. In this review, we provide an overview of the latest progress in the application of deep learning to the imaging assessment of bladder cancer. First, we review the current deep learning approaches used for bladder segmentation. We then provide examples of how deep learning helps in the diagnosis, staging, and treatment management of bladder cancer using medical images. Finally, we summarize the current limitations of deep learning and provide suggestions for future improvements.
Introduction
Bladder cancer (BCa) remains a significant global health burden, ranking as the tenth most common cancer worldwide. Currently, its diagnosis relies heavily on invasive and costly procedures such as cystoscopy and transurethral resection of bladder tumors (TURBT). While non-invasive imaging modalities like MRI, CT, and PET are available, the inherent complexity of BCa—characterized by its "hollow organ" nature where shape, position, and volume vary constantly—presents a substantial challenge for consistent manual radiological assessment. Deep learning (DL) offers a promising solution by providing automated, objective, and high-precision tools to support clinical decision-making.
The Evolution of Deep Learning in BCa Imaging
The application of deep learning to BCa imaging has evolved rapidly over the last decade, transitioning from generic convolutional neural network (CNN) architectures to highly specialized models designed for medical tasks.
As shown in Figure 1, the "Deep Learning Era" began to impact BCa imaging significantly around 2016. Since then, research has branched into three primary workflows: automated segmentation, diagnosis/staging, and treatment response assessment.
Methodological Workflows
1. Segmentation
Automated segmentation is the foundational step in BCa imaging analysis. Early CNNs have largely been superseded by the U-Net architecture, which now serves as the backbone for most modern models. Specialized variations have emerged to handle the bladder's complex morphology, such as PiPNet (Pyramid in Pyramid) for multi-scale feature extraction and CPA-Unet for fine-grained segmentation of the bladder wall and tumor regions.
2. Diagnosis and Staging
Accurate staging is critical for determining treatment pathways, particularly in differentiating between Non-Muscle-Invasive (NMIBC) and Muscle-Invasive (MIBC). The latter requires aggressive interventions like radical cystectomy. Models such as VGG16, VGG19, and ResNet18 have been successfully deployed to automate this differentiation with high sensitivity.
3. Treatment Assessment
Evaluating a patient's response to neoadjuvant chemotherapy is another vital application. Current state-of-the-art approaches often utilize hybrid Computer-Aided Decision Support Systems (CDSS). These systems combine DL-CNNs for automated feature extraction with radiomics—hand-crafted quantitative features—to provide a more comprehensive assessment than either method alone.
Key Insights and Clinical Impact
A major finding in recent literature is that geometric context is essential. Standard CNNs often struggle with the bladder's variable morphology. Models that integrate "subject-specific shape information" or Adaptive Shape Priors (ASP) significantly outperform generic models.
The clinical impact of these technologies is becoming increasingly evident. Recent 3D CNN models have achieved a Dice Similarity Coefficient (DSC) of 0.97–0.99 for bladder wall segmentation. In staging tasks, VGG16-based models have reached an AUROC of 0.997, demonstrating near-perfect classification in controlled test environments. Furthermore, studies have shown that DL-based decision support systems can significantly improve the accuracy of physicians in evaluating chemotherapy responses.
Limitations and Future Directions
Despite these advancements, several hurdles remain for widespread clinical adoption:
- Data Scarcity: Most studies rely on small, single-center datasets (ranging from 17 to 441 patients). There is a critical need for large-scale, multi-center public benchmarks to ensure model generalizability.
- Modality Expansion: While T2-weighted MRI and CT are the primary focuses today, future research should explore Diffusion-Weighted Imaging (DWI) and Dynamic Contrast Enhancement (DCE) for better physiological insights.
- Multimodal Fusion: The next frontier involves combining imaging data with genomics and pathology to create a holistic "digital twin" of the patient.
- Interpretability: Addressing the "black box" nature of DL algorithms is essential to build the trust necessary for clinical integration.
Summary updated