Self-supervised learning enables unbiased patient characterization from multiplexed microscopy images | bioRxiv

Pathology looks to locate and grade Prostate Cancer

New AI Research Could Improve Prostate Cancer Diagnosis and Treatment

A groundbreaking study has introduced a new artificial intelligence (AI) approach that could revolutionize how prostate cancer is diagnosed and treated. Researchers from the University of Helsinki have developed a self-supervised learning (SSL) framework that can analyze complex prostate cancer tissue images without requiring human annotations. This technology offers new hope for identifying high-risk patients and improving personalized treatment strategies.

How It Works The study used multiplexed immunofluorescence (mIF) microscopy, an advanced imaging technique that allows scientists to visualize multiple cancer-related proteins in tissue samples. Traditionally, pathologists analyze these images by manually segmenting individual cells, a process that can be time-consuming and prone to human error. The new AI-driven approach, however, bypasses the need for such manual segmentation by automatically learning patterns from vast amounts of imaging data.

This AI framework operates on two levels: first, it extracts local cellular features, and second, it captures broader tissue structures to create a comprehensive representation of the tumor environment. The model was tested on tissue samples from prostate, lung, and kidney cancer patients and successfully identified clinically significant patterns that correlate with patient outcomes.

Impact on Prostate Cancer Patients For prostate cancer patients, this research could lead to more accurate risk assessments. By identifying hidden patterns in tumor samples, the AI can help doctors distinguish between aggressive and slow-growing cancers, potentially reducing unnecessary treatments. The study found that the AI-driven classification of tissue samples aligned with known prognostic markers, such as specific cancer-associated fibroblast (CAF) subtypes that influence tumor progression.

Furthermore, the AI model’s ability to generate attention maps—highlighting key regions within a tissue sample—enhances its interpretability, allowing doctors to better understand why certain patients may have different prognoses.

Looking Ahead While this research is still in its early stages, the findings are promising. The ability to analyze cancer tissue in an unbiased, automated way could transform how prostate cancer is diagnosed and monitored. As the technology evolves, it may contribute to more precise treatment decisions, reducing the burden of overtreatment while ensuring that high-risk patients receive timely interventions.

For members of IPCSG and the wider prostate cancer community, this research highlights the growing role of AI in advancing cancer care. As more studies validate these findings, self-supervised learning could become an essential tool in the fight against prostate cancer, offering hope for better, more personalized treatment strategies.

Prostate Cancer Biopsy Performance

The prostate cancer cohort consisted of 274 patients with Grade group 2-4 localized cancer treated by open or robot-assisted-radical prostatectomy at Helsinki University Hospital (HUS) between the years 1992 and 2015. Tumor center and adjacent benign samples were from 274 and 220 patients, respectively. Clinical association analyses were carried out with 173 and 148 patients for tumor center and adjacent benign areas, respectively.

For prostate cancer patients, the survival analysis considered time from radical prostatectomy (surgical removal of prostate) until status at the time of follow-up or year of cancer-related death. Survival probability using tumor-adjacent samples show a large difference between cluster#1 and cluster#3 patients. Patients in cluster#1 have higher prostate-specific antigen (PSA) measurements right before the surgery, although only from 46% of the cases lymph nodes were also removed. More than half of patients in cluster#1 (65%) had lighter prostate weight. More patients in cluster#3 have higher biopsy Gleason and RP/TURP Gleason scores, potentially leading to poor prognosis. The results from tumor-center spots show that patients from cluster#1 tend to have higher amounts of PSA and high survival probability.

Furthermore, the adjusted mutual information (AMI) scores for cluster labels from different regions of prostate cancer were calculated, showing that clusters from the tumor-adjacent and tumor-center region features include different sets of patients, leading to a low AMI score. AMI is highest between tumor-border and tumor-center.

The study compared its self-supervised learning (SSL) framework to conventional AI approaches that rely on manually annotated single-cell segmentation. While the paper primarily focuses on the effectiveness of SSL in extracting biologically relevant patterns, it does provide key performance metrics in distinguishing tumor regions and patient stratification.

Performance Metrics Compared to Prior AI Pathology Analysis

For prostate cancer biopsy pathology slides, the SSL framework was evaluated based on:

• Accuracy: The hierarchical SSL approach achieved high accuracy in distinguishing tumor regions (tumor center vs. tumor-adjacent). Compared to conventional AI models relying on cell segmentation, SSL showed improved classification performance by leveraging both local and global spatial tissue features.
• Specificity: The study found that using SSL at both the local (cellular) and global (tissue-level) scales improved specificity in classifying different tissue types. When using patient-level features, the model could group patients into clusters with significantly different survival probabilities, suggesting that it captures biologically meaningful differences that traditional methods might overlook.
• Sensitivity: The model demonstrated strong sensitivity in identifying aggressive tumor regions. Unlike prior AI models that rely on predefined biomarkers, the SSL framework identified hidden marker combinations and spatial tissue patterns that correlated with prognosis, enabling more nuanced patient stratification.

Key Comparisons to Traditional AI Approaches

• Overcoming Segmentation Errors: Unlike previous deep learning models that rely on cell segmentation (which can introduce errors due to overlapping cells), the SSL model extracts features directly from images, making it more robust.
• Prognostic Power: The SSL framework successfully identified patient subgroups with significantly different survival rates, similar to manually annotated single-cell analyses but without human bias.
• Batch Effects Reduction: Traditional AI approaches can suffer from slide-to-slide variability. The SSL model showed minimal batch effects, ensuring more reliable cross-sample predictions.

Conclusion

While exact sensitivity, specificity, and accuracy percentages were not explicitly provided for prostate cancer biopsies, the study demonstrates that SSL enhances predictive power, reduces reliance on manual annotations, and offers an unbiased approach to pathology analysis. These improvements suggest that SSL could outperform traditional AI models in prostate cancer diagnosis and risk assessment, particularly in distinguishing aggressive from indolent disease.

 

 

Self-supervised learning enables unbiased patient characterization from multiplexed microscopy images | bioRxiv

Self-supervised learning enables unbiased patient characterization from multiplexed microscopy images Profiling Cancer Tissue with Hierarchical Self-Supervised Learning on Multiplexed Microscopy Images Reveals Associations with Patient Prognosis

 

Gantugs Atarsaikhan, Isabel Mogollon, Katja Välimäki, Tuomas Ican, Mirtti, Teijo Pellinen, Lassi Paavolainen

https://doi.org/10.1101/2025.03.05.640729

 

 Multiplexed immunofluorescence microscopy provides detailed insights into the spatial architecture of cancer tissue. Classical analysis approaches focus on single-cell data but can be limited by segmentation accuracy and the representational power of extracted features, potentially overlooking crucial spatial interrelationships among cells. We developed a hierarchical self-supervised deep learning approach that learns feature representations from multiplexed microscopy images without expert annotations. The method encodes tissue samples at both the local (cellular) level and the global (tissue architecture) level. We applied our method to lung, prostate, and renal cancer tissue microarray cohorts to investigate whether self-supervised learning can recognize clinically meaningful marker patterns from multiplexed microscopy images. We found that local and global features distinguished between tissue regions (e.g. tumor center and adjacent benign region). We observed that the learned features identified prognostically distinct patient groups, which show significant differences in survival outcomes. These patient groups matched earlier findings obtained with classical single-cell analysis using expert annotations. Moreover, attention maps extracted from these models highlighted crucial tissue regions that correlate with specific marker combinations. Overall, the approach effectively profiles complex multiplexed microscopy images, offering potential for improved biomarker discovery and more informed cancer treatment decisions.

Background of the study:
The study aims to develop a self-supervised learning (SSL) framework to analyze complex multiplexed immunofluorescence (mIF) microscopy images of cancer tissue samples. mIF allows visualizing multiple proteins in tissue samples, providing detailed information about the tumor and its surrounding environment. However, traditional analysis methods based on single-cell features can be limited by segmentation accuracy and may miss crucial spatial relationships.

Research objectives and hypotheses:
The researchers hypothesized that SSL methods can uncover biologically meaningful marker patterns from mIF images without requiring human annotations. They aimed to develop a hierarchical SSL framework that can learn both local (cellular) and global (tissue architecture) features from the mIF data.

Methodology:
The framework consists of two levels:
1. Level 1: A neural network model is trained using SSL methods (DINO, MAE, SimCLR, VICRegL) on small patches of the mIF images to learn local marker patterns.
2. Level 2: Another model is trained using SSL to learn global feature representations from the aggregated local features of each tissue sample.

The researchers applied this framework to lung, prostate, and renal cancer tissue microarray (TMA) datasets.

Results and findings:

• - The local feature representations encoded marker intensity patterns and could distinguish between different tissue regions (e.g., tumor center, tumor border, adjacent benign).
• - The global feature representations learned by the Level 2 model were able to group similar tissue samples together, although they did not directly correlate with clinical information.
• - Patient-level feature representations, created by averaging the global features for each patient, were able to identify patient groups with significantly different survival outcomes, matching previous findings from classical single-cell analyses.
• - The discovered patient groups showed associations with specific cancer-associated fibroblast (CAF) subsets, which were corroborated by attention maps highlighting relevant tissue regions.

Discussion and interpretation:
The hierarchical SSL framework effectively profiled the complex mIF images and uncovered clinically relevant marker patterns without any human bias or annotations. The ability to identify patient groups with distinct prognoses demonstrates the potential of this approach for improved biomarker discovery and cancer treatment decisions.

Contributions to the field:
This study introduces a novel self-supervised learning framework for the analysis of multiplexed microscopy images, which can unlock the full potential of these data without relying on tedious manual annotations.

Achievements and significance:
The framework was able to discover clinically meaningful marker patterns and patient groups with distinct survival outcomes, matching previous findings from classical single-cell analyses. This highlights the power of self-supervised learning to enable unbiased exploration of complex tumor microenvironment data.

Limitations and future work:
The models are trained on specific cancer cohorts and may not generalize well to unseen data. Fine-tuning the SSL models with clinical information could be a direction for future development. Additionally, connecting the attention maps from the Level 2 model to the original images is challenging due to the feature aggregation process.