Clinical Missingness-Aware mmFormer for Incomplete Multimodal Brain Tumor Segmentation
IOSR Journal of Computer Engineering · 2026-03 · 28 (3) · 55-64
DOI: 10.9790/0661-2803025564 ↗Kurz gefasst
Fachartikel in der Zeitschrift IOSR Journal of Computer Engineering, 2026-03. Verfasst von Thanh Trung Giang.
Abstract (Originalsprache)
Multimodal magnetic resonance imaging plays an important role in brain tumor segmentation because different MRI sequences provide complementary information about tumor subregions; however, most existing multimodal segmentation methods assume that all modalities are available and reliable during inference, an assumption that is often unrealistic in clinical practice where MRI modalities may be missing or degraded by noise, motion artifacts, blur, or low signal quality. To address this limitation, this study proposes Clinical Missingness-Aware mmFormer (CMA-mmFormer) for incomplete multimodal brain tumor segmentation. The proposed framework extends mmFormer by introducing a clinical missingness simulation protocol, a Missingness Encoder, and a Reliability Gating mechanism. The model was evaluated on the BraTS 2018 dataset using four MRI modalities (FLAIR, T1ce, T1, and T2) under multiple missingness scenarios, including complete modalities, missing T1ce, FLAIR + T2 limited protocol, single-modality input, and qualitydegraded modalities. Segmentation performance was assessed using Dice score and HD95, while prediction reliability was evaluated using expected calibration error, Brier score, negative log-likelihood, and riskcoverage analysis. Experimental results showed that CMA-mmFormer outperformed the original mmFormer baseline, improving WT Dice from 84.20% to 87.10%, TC Dice from 76.80% to 81.50%, and ET Dice from 72.40% to 78.30%, while also reducing HD95 values across all tumor subregions. Under clinical missingness scenarios, CMA-mmFormer demonstrated consistent improvements, particularly in missing T1ce and qualitydegraded settings. In addition, the model achieved better calibration, reducing ECE from 0.082 to 0.049 and AURC from 0.214 to 0.151. These findings indicate that CMA-mmFormer improves incomplete multimodal brain tumor segmentation by explicitly modeling both modality availability and modality reliability, and that the proposed Missingness Encoder and Reliability Gating mechanism enhance segmentation robustness and prediction reliability under clinically realistic missing and degraded MRI conditions, highlighting the potential of clinical missingness-aware modeling for trustworthy multimodal medical image analysis.
Zitieren
APA
Giang, T. T. (2026). Clinical Missingness-Aware mmFormer for Incomplete Multimodal Brain Tumor Segmentation. IOSR Journal of Computer Engineering, 28(3), 55-64. https://doi.org/10.9790/0661-2803025564
Vancouver
Giang TT. Clinical Missingness-Aware mmFormer for Incomplete Multimodal Brain Tumor Segmentation. IOSR Journal of Computer Engineering. 2026;28(3):55-64. doi:10.9790/0661-2803025564
BibTeX
@article{Giang2026,
author = {Giang, Thanh Trung},
title = {Clinical Missingness-Aware mmFormer for Incomplete Multimodal Brain Tumor Segmentation},
journal = {IOSR Journal of Computer Engineering},
year = {2026},
volume = {28},
number = {3},
pages = {55--64},
doi = {10.9790/0661-2803025564}
}
EUVIMED