REASON: Probability map-guided dual-branch fusion framework for gastric content assessment

TL;DR

REASON tackles automated gastric content assessment from ultrasound to reduce aspiration risk by introducing a two-stage framework: probability map guidance (PMG) to focus on gastric regions via semi-supervised segmentation, and a dual-branch fusion classifier (DBFC) to integrate information from right-lateral decubitus (RLD) and supine (SUP) views. The method leverages a mean-teacher training regime with Bidirectional Copy-Paste to generate informative probability maps $p\in[0,1]^{C_0\times H\times W}$ and enhances multi-view inputs with $x'=(1-\gamma)x+\gamma x\odot p$, followed by logits fusion $y_f=\beta y_r+(1-\beta)y_s$ (with $\beta=0.7$). Experimental results on a self-collected dataset of $2{,}174$ images from $364$ patients show state-of-the-art accuracy (Acc $=82.15\%$) and F1 (≈$82.1\%$), with ablations confirming the complementary gains from PMG and DBFC and identifying DenseNet121 as the strongest backbone. The work demonstrates robust, efficient gastric content classification with potential for automated preoperative risk stratification and clinical deployment, while highlighting avenues for cross-center generalization.

Abstract

Accurate assessment of gastric content from ultrasound is critical for stratifying aspiration risk at induction of general anesthesia. However, traditional methods rely on manual tracing of gastric antra and empirical formulas, which face significant limitations in both efficiency and accuracy. To address these challenges, a novel two-stage probability map-guided dual-branch fusion framework (REASON) for gastric content assessment is proposed. In stage 1, a segmentation model generates probability maps that suppress artifacts and highlight gastric anatomy. In stage 2, a dual-branch classifier fuses information from two standard views, right lateral decubitus (RLD) and supine (SUP), to improve the discrimination of learned features. Experimental results on a self-collected dataset demonstrate that the proposed framework outperforms current state-of-the-art approaches by a significant margin. This framework shows great promise for automated preoperative aspiration risk assessment, offering a more robust, efficient, and accurate solution for clinical practice.

Figures (10)

• Figure 1: Comparison between the conventional clinical workflow and the proposed method. (a) The conventional clinical workflow, which relies on empirical formulas, is labor-intensive and suffers from limited generalizability. (b) In contrast, the proposed learning-based approach offers superior efficiency and robustness.
• Figure 2: Overall framework of REASON. It consists of two stages. In stage 1, probability maps generated by the segmentation model are used to highlight anatomically relevant gastric regions and suppress artifacts. In stage 2, the dual-branch fusion classifier integrates complementary spatial and contextual features from the RLD and SUP views to enable robust gastric content assessment.
• Figure 3: Bidirectional copy-paste strategy. BCP generates composite images that combine regions of strong supervision (from ground truth) with weak supervision (from pseudo-labels), allowing the former to effectively regularize and improve the latter.
• Figure 4: Dual-branch fusion classifier. It integrates complementary spatial and contextual information from SUP and RLD views through a weighted combination of logits.
• Figure 5: Confusion matrices of several representative methods. (a) DenseNet121. (b) EfficientNet-B0. (c) MobileNet-V3. (d) SoftAug. (e) MedMamba. (f) REASON (Ours). Single view-based methods (a)-(e) sum ten confusion matrices (5 folds $\times$ 2 views), while our dual view-based REASON (f) sums five confusion matrices (5 folds).