Brain tumor segmentation using deep learning: high performance with minimized MRI data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Manual brain tumor segmentation from MRI is labor-intensive and time-consuming.
• Deep learning (DL) offers a potential solution to automate and optimize this process.
• Minimizing the number of required MRI sequences can enhance DL model generalizability and clinical adoption.

Purpose of the Study:

• To optimize deep learning-based brain tumor segmentation by minimizing the number of MRI sequences.
• To evaluate the performance of a 3D U-Net model using different combinations of MRI sequences for glioma segmentation.
• To compare the segmentation accuracy of models trained on T1C-only, FLAIR-only, T1C + FLAIR, and T1 + T2 + T1C + FLAIR sequences.

Main Methods:

• Trained a 3D U-Net deep learning model on the 2018 MICCAI BraTS dataset.
• Focused on sub-segmenting enhancing tumor (ET) and tumor core (TC).
• Evaluated model performance using 5-fold cross-validation and a separate test dataset (358 samples), comparing four MRI sequence combinations.

Main Results:

• The T1C + FLAIR sequence combination achieved Dice scores comparable to or exceeding the full four-sequence set in both cross-validation and test datasets for ET and TC segmentation.
• T1C-only also showed strong performance for TC segmentation.
• Across all configurations, specificity remained high (≥0.958), with T1C + FLAIR demonstrating superior performance in ET delineation.

Conclusions:

• Deep learning models can achieve high accuracy in brain tumor segmentation using only two MRI sequences (T1C + FLAIR).
• Reducing reliance on multiple sequences can improve DL model generalizability and facilitate wider dissemination in clinical and research settings.
• This approach has the potential to significantly reduce the manual labor involved in brain tumor segmentation.