Artikel
3D MRI segmentation of glioma – research datasets vs. routine clinical data
3D MRT Segmentierung von Glioma – Forschungs-Datensätze vs. Daten aus dem klinischen Alltag
Suche in Medline nach
Autoren
Veröffentlicht: | 4. Juni 2021 |
---|
Gliederung
Text
Objective: Automated segmentation of brain tumors in 3D MR images has seen significant advances in recent years. State-of-the-art algorithms rely on 3D MRI sequences to achieve best results: pre-contrast T1- weighted (T1), post-contrast T1-weighted (T1CE), T2-weighted (T2), and T2 Fluid Attenuated Inversion Recovery (FLAIR). In clinical routine, T2 and FLAIR sequences are typically only acquired in 2D, not suitable as input for current algorithms. We aim to quantify the loss of information when using only the two 3D T1 sequences instead of all four 3D sequences.
Methods: 3D MR images and ground-truth segmentations of n = 369 patients (293 subjects are HGG, while 76 are LGG cases, mean age 61 (minimum=18 and maximum=86)) were obtained from the Brain Tumor Segmentation (BraTS) Challenge 2020. A set containing ST1±CE:= (T1, T1CE), omitting T2 and FLAIR, and SFULL:= (T1, T1CE, T2, FLAIR), containing all sequences were defined. An encoder-decoder based CNN architecture with an asymmetrically larger encoder (VAE-RES-NET) was adapted for tumor segmentation using a batch size of four, a cropsize of 160*192*128, trained for 300 epochs with an early stopping of 100. The net architecture is shown in Figure 1 [Fig. 1]. We trained two VAE-RES-NET one for each sequence set shown in Figure 2 [Fig. 2]. The algorithm was trained (80% of the dataset) to automatically segment the following three labels: the contrast-enhancing tumor (ET), the peritumoral edema (ED) and the necrotic and non-enhancing tumor core (NCR/NET). The accuracy was evaluated using the Dice coefficient averaged over the validation set (20% of the dataset).
Results: The average Dice score for the label NCR/NET was 0.674 for set ST1±CE and 0.685 for set SFULL. For label ET, the Dice coefficient was 0.760 and 0.764, respectively. A marked difference was found for the label ED, with 0.652 for set ST1±CE and 0.800 for set SFULL, a difference of 0.147.
Conclusion: Using incomplete information, i.e. omitting T2 and FLAIR, NCR/NET and ET were segmented with an accuracy comparable to the full set of sequences, while ED segmentation accuracy declined. Even though epitumoral edema is mainly encoded in T2 and FLAIR imaging, reasonable accuracy was obtained using incomplete information. Nevertheless, since in clinical routine T2 and FLAIR sequences are often only acquired in 2D, further work should focus on including these in the modelling to boost ED segmentation accuracy.