A deep learning-based automatic segmentation model for diffuse midline glioma with H3K27M alteration.

Area of Science:

• Neuro-oncology
• Medical imaging
• Artificial intelligence in medicine

Background:

• Diffuse midline glioma (DMG) with H3K27M alteration is a fatal brain tumor.
• Radiotherapy is a preferred treatment, but precise target delineation is difficult.
• Accurate radiotherapy planning requires robust tumor segmentation.

Purpose of the Study:

• To develop and validate a deep learning model for automatic radiotherapy target delineation in diffuse midline glioma (DMG) with H3K27M alteration.
• To improve the accuracy and efficiency of radiotherapy planning for DMG patients.

Main Methods:

• Utilized contrast-enhanced T1-weighted (T1C), T2-weighted, and T2 fluid attenuated inversion recovery (T2-Flair) MRI sequences.
• Developed a generative adversarial network (GAN)-based model incorporating spatial channel attention and multi-scale feature extraction.
• Trained and tested the model on data from 116 and 26 patients, respectively, across two medical centers.

Main Results:

• The T2 sequence model achieved the highest Dice Similarity Coefficient (DSC) of 0.916 in the training set.
• In the test set, DSC values were 0.750 (T1C), 0.872 (T2), and 0.862 (T2-Flair), indicating strong generalizability.
• The model demonstrated excellent automatic segmentation accuracy for DMG with H3K27M alteration.

Conclusions:

• Successfully developed the first GAN-based automatic segmentation model for DMG with H3K27M alteration.
• The model exhibits excellent accuracy and generalizability for radiotherapy target delineation.
• This deep learning approach offers a promising tool for improving radiotherapy in DMG patients.