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

Background: Diffuse midline glioma (DMG) is a fatal tumor that emerges in the brainstem and thalamus. Compared with microsurgery and chemotherapy, radiotherapy is currently regarded as a safer and more effective treatment option. However, mapping radiotherapy target on brain stem is extremely demanding. In the study, we build a deep learning-based diffuse midline glioma with H3K27M alteration radiotherapy target area automatic delineation model. Methods: We collected contrast-enhanced T1-weighted (T1C), T2-weighted, and T2 fluid attenuated inversion recovery (T2-Flair) sequences from patients with DMG and H3K27M alteration from two medical centers to train and test the model. Based on the framework of generative adversarial networks (GANs), we integrated spatial channel attention mechanism and multi-scale feature extraction according to the characteristics of tumor location in the midline region and diverse morphological changes. Results: The training and test sets included 116 and 26 patients, respectively. In the training set, the segmentation performance was best for the T2 sequence model, with a Dice similarity coefficient (DSC) of 0.916, followed by the T2-Flair sequence model, with a DSC of 0.893; and the T1ce sequence model had the lowest segmentation accuracy, with a DSC of 0.802. In the test set, the DSC values for the T1C, T2, and T2-Flair sequence models were 0.750, 0.872, and 0.862, respectively, demonstrating the strong generalizability of the model. Conclusions: We developed DMG with H3K27M alteration automatic segmentation model based on GANs for the first time. It shows excellent automatic segmentation accuracy and generalizability.