Oral Presentation 36th TROG Cancer Research Annual Scientific Meeting 2024

Machine learning simulation of targeted ablation for oligoprogressive metastatic prostate cancer (#31)

Mikaela Dell'oro 1 , Timothy G Perk 2 , Ojaswita Lokre 2 , Colin Tang 3 , Martin A Ebert 1 3 4 5 , Roslyn J Francis 1 6
  1. Australian Centre for Quantitative Imaging, School of Medicine, The University of Western Australia, Perth, WA 6009, Australia
  2. AIQ Solutions, Madison, WI 53717, USA
  3. Department of Radiation Oncology, Sir Charles Gairdner Hospital, Perth, WA 6009, Australia
  4. School of Physics, Mathematics and Computing, The University Of Western Australia, Perth, WA 6009, Australia
  5. School of Medicine and Public Health, University of Wisconsin, Madison, WI 53705, USA
  6. Department of Nuclear Medicine, Sir Charles Gairdner Hospital, Perth, WA 6009, Australia

Background:
Treatment response heterogeneity is common in oncology patients. Oligoprogression is when a subset of progressing lesions drives treatment resistance. Automated detection and machine learning models have the potential to provide assistive tools for identification of oligoprogressive lesions that may be suitable for targeted ablation.

 

Aim:

This investigation simulated the impact of targeted ablation of oligoprogressive lesions in metastatic prostate cancer (mPC) patients using [68Ga]Ga-PSMA PET/CT scans.

 

Methods:

[68Ga]Ga-PSMA PET/CT imaging and overall survival (OS) data from 195 patients with mPC biochemical relapse were assessed. PSMA-positive lesions were defined by a Nuclear Medicine physician at baseline (BL) and 6-month follow-up (FU). TRAQinform IQ technology (AIQ Solutions) was used to track individual lesions across timepoints and categorise as new, increasing, stable, decreasing, or disappeared based on changes in SUVtotal. Imaging features extracted from each patient were used to train a random forest survival model (evaluated using c-index). For each patient, up to five increasing or new lesions were simulated as ablated. This was achieved through removal of the lesions, simulating ablation by removing the lesion from the second scan and recomputing all imaging features post-simulation. The TRAQinform Profile was used to generate a risk score of reduction in probability of death for each patient with and without simulated radiation therapy.

 

Results:

Of 195 patients, 1,233 lesions were identified at BL, and 1,605 were identified at FU. The TRAQinform Profile score was a strong predictor of OS with a c-index of 0.83 across all patients. After the simulated removal by ablation 79/195 patients had a reduction in probability of death within 5 years (mean: 22%, range: 10-50%); 42/79 of which had a single lesion ablated.

 

Conclusions:

TRAQinform Profile identified lesions for targeted radiation therapy that may reduce probability of death by an average of 22% within 5 years in a simulated model.