# Predicting prostate cancer recurrence using an atlas‐based tumor control probability model

## TL;DR

This study shows that an atlas-based tumor control probability model can predict prostate cancer recurrence by integrating patient-specific data and histopathology reports.

## Contribution

The novel contribution is the use of segment-wise adjustments to TCP model parameters based on histopathology and biological atlases for improved recurrence prediction.

## Key findings

- Combining cell density and GS-dependent α/β adjustments predicted recurrence in all nine patients with significant TCP reductions.
- Lower TCP regions aligned with relapsed tumor sites in 78% of patients, supported by statistical analysis of TCP differences.
- GP-dependent α adjustments failed to predict recurrence, while cell density alone showed moderate performance.

## Abstract

Recurrence following prostate cancer (PCa) radiation therapy (RT) remains a persistent challenge. Although dose escalation can improve tumor control, it often results in increased toxicity. With an understanding of tumor heterogeneity, identification of radioresistant tumor subvolumes at risk of low tumor control probability (TCP) could provide an opportunity for personalized dose prescription to reduce risk of treatment failure without the increased risk of toxicity.

The aim of this study was to evaluate the efficacy of an atlas‐based tumor control probability model in predicting prostate cancer recurrence by retrospectively integrating patient‐specific primary radiotherapy and histopathology‐informed data. A segment‐wise adjustment to TCP model parameters was investigated for enhancing recurrence prediction in a patient cohort with biopsy‐confirmed local recurrence following definitive RT.

Nine patients with biopsy proven local recurrence were selected from an ethics‐approved study (NCT03073278) based on the availability of histopathology reports, dose‐fractionation schedules and treatment planning data from their primary RT. Two previously‐reported population‐based biological atlases, one comprising a cell density data (CD‐atlas) and the other tumor probability data (TP‐atlas), were deformably registered to the prostate contour of each patient. Histopathology reports were retrieved for each patient, and the registered prostate atlases were anatomically segmented based on individual histopathology findings. Radiosensitivity parameters were derived from a separate patient cohort's histology dataset using a numerical optimization method, generating a single PCa grade‐independent α/β ratio, four Gleason Pattern (GP)‐dependent α parameters, and nine Gleason Score (GS)‐dependent α/β ratios. Three parameter adjustment approaches—cell density alone, cell density with GP‐dependent α, and cell density with GS‐dependent α/β, were evaluated and compared to a baseline model without adjustments. Changes in overall TCP values resulting from the adjustments were analyzed, and recurrent gross tumor volume (GTV) contours were overlaid on the TCP maps to evaluate their alignment with regions of lower TCP, assessing the model's ability to enhance recurrence prediction.

The approach combining segment‐wise cell density and GS‐dependent α/β adjustments showed superior predictive capability, with all nine patients (100%) exhibiting a significant (p = 0.004) reduction in overall TCP values and seven patients (78%) showing alignment of lower TCP regions with relapsed tumor sites. This was further supported by voxel‐level histogram analysis and statistically significant volume‐weighted TCP differences between GTV and nonGTV regions (Wilcoxon signed‐rank test, p = 0.003). In contrast, GP‐dependent α adjustments alongside cell density failed to predict recurrence, while cell density adjustments alone yielded moderate performance. Additionally, the generated single α/β ratio and GS‐dependent α/β ratios were consistent with the lower α/β ratios typically associated with PCa.

The atlas‐based TCP model, enhanced with patient‐specific histopathology report data, demonstrated promising capabilities in predicting PCa recurrence. This approach has the potential to support personalized treatment planning by enabling optimization of the distribution of a specific integral dose to minimize recurrence risk.