Table 4 Performance comparison with state-of-the-art architectures in the ISBI challenge.

Approach	Modalities	CNN type	DSC	PPV	TPR	LFPR	LTPR	Submission score
2.5D Tiramisu¹⁷	FLAIR, T1w, T2w, PD	2.5D	0.64	0.91	0.53	0.12	0.52	93.358
ALL-NET¹⁸	FLAIR, T1w, T2w	3D	0.63	0.91	–	0.12	0.533	93.32
nnUNet¹⁹	FLAIR, T1w, T2w, PD	3D cascade	0.69	0.85	0.61	0.17	0.55	93.09
nnUNet²⁰	FLAIR, T1w, T2w, PD	3D	0.68	0.86	0.60	0.16	0.54	93.03
DeepLesionBrain²¹	FLAIR & T1w	3D	0.65	0.89	0.55	0.13	0.49	92.85
Multi-branch U-Net (proposed)	FLAIR, T1w, T2w	2D	0.64	0.85	0.56	0.20	0.55	92.661
IMAGINE²²	FLAIR, T1w, T2w, PD	3D	0.58	0.92	0.46	0.09	0.41	92.49
Self-adaptive network ²⁴	FLAIR. T1w, T2w, PD	3D	0.68	0.78	0.65	0.27	0.60	92.41
Multi-branch ResNet ²⁵	FLAIR, T1w, T2w	2D	0.61	0.90	0.49	0.14	0.41	92.12
Attention-Based CNN ²⁶	FLAIR, T1w	3D	0.64	–	–	0.39	0.45	–

This table compares the performance metrics of state-of-the-art published architectures with the proposed architecture. The proposed method lands on rank 6, where it is able to outperform all 2D approaches and even two of the state-of-the-art 3D approaches, regarding the submission score. The best and second-best results are written in bold and italic, respectively.

Quick links

Search