Related Experiment Video
Updated: Jul 3, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
A Systematic Evaluation of Image Preprocessing in Deep Learning Detection and Segmentation of Intracranial Metastatic
Peter I Kamel1, Ahmed Naeem2, Komal Shah2
1From the Department of Neuroradiology (P.I.K., A.N., K.S., M.W.), Division of Diagnostic Imaging, MD Anderson Cancer Center, Houston, TX and Department of Radiology (M.W.), The University of Texas Medical Branch, Galveston, TX. peterkamelmd.correspondence@gmail.com.
Background And Purpose:
Image preprocessing is an essential, though often overlooked, part of machine learning, and it is unclear how preprocessing techniques affect metastatic disease segmentation. This is particularly true given the differences between segmentation of primary brain tumors and the detection and segmentation of small metastases. The purpose of this work is to systematically evaluate how preprocessing techniques affect the detection and segmentation of intracranial metastases.
Materials And Methods:
Using nnU-Net, we evaluated preprocessing techniques on a training data set of 1,034 MRIs and testing set of 256 MRIs. We examined the effects of label preparation, bias correction, normalization, registration, and denoising, as well as the contribution of each MRI sequence. Model performance was compared to the baseline model using paired t-tests, evaluating global Dice scores, lesion detection rate, and per-lesion and per-exam Dice score, surface Dice score, and 95% Hausdorff distance, stratified by lesion volume.
Results:
Baseline model performance demonstrated a global Dice score of 0.72 ± 0.28 and a lesion detection rate of 70.8% (1,328/1,877), strongly dependent on lesion size. Including all labels in training, as opposed to only tumor core, resulted in a significant decrease in detection rate to 68.8% (1,292/1,877) (p < 0.001), as well as reduced per-lesion segmentation particularly for smaller lesions. Bias correction and denoising demonstrated no significant effect on performance. Pre-normalization of data led to worse segmentation performance. Image registration reduced both lesion detection and segmentation. Training with fewer MRI sequences, in several cases, improved segmentation.
Conclusions:
Our findings demonstrate that certain preprocessing steps have a significant impact on model performance while others have negligible effects and may not be necessary. With nnU-Net, we observed highest detection and segmentation scores by reducing training labels and MRI sequences to those focused on tumor core, with additional, less relevant data, reducing performance. Several preprocessing steps showed no measurable benefit and image registration may be detrimental to the detection and segmentation of small metastases.