---
title: >-
  Princess Margaret Cancer Centre / University Health Network: ML-assisted
  treatment-plan generation and blinded physician selection
slug: prostate-ml-radiotherapy-planning
stable_id: b8243d473b7bd45c
company: Princess Margaret Cancer Centre / University Health Network
function_code: clinical_ops
pattern_codes:
  - continuous_decisioning
  - creator_to_judge
evidence_strength: verified
publication_tier: showcase
freshness: current
reviewed_at: '2026-08-23'
updated_at: '2026-08-22'
source_quality_summary: 2 peer reviewed; publication outcomes are verified.
caveat_summary: >-
  Single-center prospective deployment; 61% of ML plans were selected prospectively
  versus 83% in simulation.
collections: []
bundle_version: 1.0.0
bundle_fingerprint: sha256:c23c6cc2b88153d008ea8fda928f632ce0011fc2d4c5036672a16e5d895bab93
canonical_url: https://brianletort.ai/transformations/prostate-ml-radiotherapy-planning
---

# Princess Margaret Cancer Centre / University Health Network: ML-assisted treatment-plan generation and blinded physician selection

A machine can generate a candidate plan, but physicians and peer review select the plan delivered.

Function: Clinical operations. Patterns: Continuous decisioning; Creator to judge. Evidence: verified.

Freshness: current. Reviewed: 2026-08-23. Updated: 2026-08-22.


Source quality: 2 peer reviewed; publication outcomes are verified.

## Before

1. **Medical dosimetrist / physicist** — Manually specifies objectives and iteratively tunes a prostate radiotherapy plan. (control: Standard planning system, dose constraints, and peer review.)
2. **Treating physician** — Reviews the human-generated plan against standardized criteria. (control: Physician and peer-review approval; median end-to-end baseline 118 hours.)

## After

1. **Random-forest planning system and human planner** — Generates an ML plan in parallel with the conventional human plan. (control: Same clinical criteria and standardized peer review; two independent candidate paths retained.)
2. **Treating physician** — Blindly compares candidates and selects the clinically acceptable plan for delivery. (control: Physician owns acceptance; unacceptable ML plans are discarded for human plans.)

## Decision rights

Treating physicians and peer review retain final plan selection and delivery authority.

## Exception path

An unacceptable ML plan is rejected in favor of the human-generated plan.

## Outcomes

- **End-to-end planning time** (verified): Median 118 hours → Median 47 hours; 60.1% reduction. Single-center prospective deployment; 61% of ML plans were selected prospectively versus 83% in simulation.

## Executive lesson

Parallel human and machine plans can expose real adoption gaps that retrospective accuracy misses.

## Anti-pattern

Do not publish simulated acceptance as the prospective adoption rate.

## Questions for leaders

- Where is the operating threshold set and who can override it?
- What measured result would trigger rollback or retraining?
- Which residual decisions must remain human-owned?

## Sources

- [Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer](https://pubmed.ncbi.nlm.nih.gov/34083812/) — Nature Medicine / PubMed
- [Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer](https://www.nature.com/articles/s41591-021-01359-w) — Nature Medicine