Training With Local Data Remains Important for Deep Learning MRI Prostate Cancer Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Domain shift negatively impacts AI model performance in medical imaging.
• Previous studies on domain shift for MRI prostate cancer segmentation used limited cohorts.
• Large-scale external datasets may not fully capture local data characteristics.

Purpose of the Study:

• To assess if AI models trained on local MRI data outperform those trained on external data for prostate cancer segmentation.
• To evaluate the impact of cohort size (>1000 exams) on domain shift effects.
• To determine the necessity of local data for robust segmentation models.

Main Methods:

• Simulated a multi-institutional consortium using public (PICAI) and local MRI datasets.
• Trained nnUNet-v2 models on combined (CENTRAL-TRAIN), external (PICAI-TRAIN), and local (LOCAL-TRAIN) data.
• Evaluated model accuracy using the PICAI Score on a local test set and tested for significance via bootstrapping.

Main Results:

• A small fraction (22%) of local training data matched external training performance.
• Models trained on combined or local data (PICAI Scores [95% CI] 65 [58-71] and 66 [60-72]) outperformed the external-only model (58 [51-64], P < .002).
• Reducing training set size did not change these performance trends.

Conclusions:

• Domain shift significantly limits MRI prostate cancer segmentation performance, even with over 1000 external exams.
• Local MRI data is crucial for achieving optimal performance in prostate cancer segmentation models at scale.
• Incorporating local data is paramount for developing reliable AI tools in this domain.