Locbam Advances 3D Segmentation, Demonstrating Improvements on BTCV, AMOS22 and KiTS23

Researchers are tackling a critical challenge in 3D medical image segmentation: efficiently processing large datasets without sacrificing accuracy. Donnate Hooft, Stefan M Fischer, and Cosmin Bercea, from the School of Computation, Information and Technology at Technical University Munich, alongside Jan C Peeken and Julia A Schnabel, present LocBAM, a new method which integrates spatial information often lost in conventional patch-based approaches. This innovation is significant because it stabilises training and markedly improves segmentation performance, especially when dealing with limited data , a common issue in medical imaging. Their findings, demonstrated on challenging datasets like BTCV, AMOS22, and KiTS23, also prove LocBAM surpasses existing coordinate encoding techniques, offering a substantial advancement in the field.

Postprocessing improved performance to 93.37% (largest component filtering) and 93.65% (atlas masking). Location-aware models outperformed both approaches, with CoordConv reaching 93.90%. Class-wise analysis (Fig0.4% even at 100% shifts, whereas CoordConv dropped to 7.67% Dice on the KiTS23 dataset and +0.17%), it resulted in substantial degradation on BTCV (-4.26%) and KiTS23 (-0.26%)0.66%, 87.44%, and 81. Using BTCV, the team examined spacings of 3.0, 0.75, 0.75, isotropic 1mm, and a low-resolution spacing of 2.0, 1.5, 1.5, training with both large (1283) and small (323) patches0.37%) and atlas masking (93.65%), achieving a mean Dice score of 93.90% on BTCV in a low-resolution, small-patch setting0.4% even with 100% axial shifts, significantly outperforming CoordConv, which dropped to 7. The authors acknowledge that their work currently focuses on axial location encoding, and future research could extend this to include coronal and sagittal axes for more refined anatomical positioning. They also suggest that data augmentation of location signals could further enhance robustness by accounting for inter-patient variability and positional uncertainty, anticipating that attention-based spatial integration will become increasingly important as imaging resolutions continue to rise and memory constraints become more challenging.