This study applied automated, quantitative assessment of collagen and extracellular matrix (ECM) architecture in diagnostic prostate biopsies to identify features associated with recurrence after radiotherapy for localised prostate cancer. The analysis used picrosirius red staining and a broad set of image-derived metrics to characterise fibre architecture, inter-fibre gaps and ECM texture, then related these parameters to tumour status, Gleason grade patterns and subsequent recurrence.
The cohort comprised 422 patients recruited to the CHHiP (Conventional or Hypofractionated High Dose Intensity Modulated Radiotherapy in Prostate Cancer) trial. Patients were matched 1:1 for recurrence status. Diagnostic prostate biopsy sections were stained with picrosirius red to visualise collagen and ECM structure for automated image analysis. The source reports that appropriate ethical approvals and patient consents were obtained; the work is presented as a preprint and has not been peer reviewed.
From the stained biopsy images the investigators derived 378 quantitative ECM parameters. These parameters comprehensively described aspects of the ECM including fibre architecture, the size and shape of gaps between fibres, and measures of ECM texture and homogeneity. Parameters were computed at the level of individual image tiles and aggregated to the patient level for further modelling.
Machine learning models were trained to relate ECM parameters to three classification problems: distinguishing tumour versus normal prostate, differentiating Gleason pattern 4 versus pattern 3, and predicting recurrence after radiotherapy. Models were applied both at the tile level (individual image regions) and at the patient level. The authors used Shapley analysis to interpret model outputs and identify which ECM features contributed most to predictions. Specific training details, model architectures, and quantitative performance metrics (for example, accuracy, AUC, or calibration) were not reported in the source.
Shapley value analysis was used to interpret the relative importance of ECM features for the models. This approach enabled the investigators to define interpretable ECM signatures associated with recurrence. The source describes how Shapley analysis highlighted patterns in gap geometry and texture as drivers of recurrence prediction, without reporting exact Shapley value magnitudes or per-feature statistics in the summary provided.
ECM patterns associated with recurrence were characterised by: smaller and more elongated gaps between collagen fibres; finer-grained ECM texture; and lower ECM homogeneity compared with regions associated with lower recurrence probability. These specific textural and gap-size features drove the recurrence signatures identified by the interpretative analysis.
The ECM signatures predictive of recurrence were enriched among patients with Gleason 4+3 disease compared with those with Gleason 3+4. The analysis therefore suggests a relationship between adverse ECM architecture and higher-grade Gleason patterns, though the source does not supply the numerical enrichment values or statistical effect sizes in the abstract.
The findings indicate that automated quantitative analysis of collagen and ECM architecture can provide information relevant to the probability of prostate cancer recurrence after radiotherapy. Features related to ECM gap size and texture appear particularly informative. The authors propose that such ECM-derived parameters could augment risk stratification in localised prostate cancer, potentially informing treatment decision-making; however, the article is a preprint, and translation to clinical practice would require peer review, external validation and prospective evaluation, details of which were not reported in the source.
The source material is an abstract and preprint without peer review. Important methodological and performance details that were not reported in the provided text include specific machine learning model types and hyperparameters, quantitative performance metrics (for example sensitivity, specificity, AUC), internal or external validation results, and how patient- or tumour-level aggregation was performed in detail. Implementation considerations for clinical workflows and prospective validation plans were not described in the source.