Systematic Evaluation: Preprocessing for Deep Learning Detection of Brain Metastases

For deep learning detection of brain metastases, registration and pre-normalization reduced detection; bias correction/denoising had no effect. Tumor-core labels and fewer MRI sequences improved performance. Including all labels decreased detection to 68.8% (p<0.001) vs tumor-co…

• Retrospective evaluation on 1,034 training and 256 test MRIs (1,877 lesions) using nnU-Net.
• Lesion size strongly influenced detection; training with fewer sequences improved segmentation in several cases.
• Single dataset, no external validation; findings specific to nnU-Net architecture.

RadPigeon summaries are original and for information only. They are not clinical advice.