バナナ0.9：低リソースの胃がんスクリーニングのためのオープンソースで再現可能な医用画像処理システム

Gastric cancer remains a major global health burden, particularly in East Asia, yet early-detection programs are often limited by computational constraints, variable imaging quality, and uneven resource availability across clinical settings. Existing AI models for CT analysis frequently require GPU-accelerated infrastructure and offer limited transparency or reproducibility, reducing their suitability for deployment in low-resource hospitals. To address these gaps, we developed and publicly released Banana 0.9, an open-source, CPU-based medical imaging framework intended to support fully reproducible, CT-based gastric cancer screening workflows. Banana 0.9 serves as a proof-of-concept milestone toward a broader, cross-cancer screening platform emphasizing interpretability, accessibility, and transparent methodology.

Banana 0.9 was implemented as a modular, GPU-free CT imaging pipeline using deterministic Hounsfield-unit (HU) rules for organ and region-of-interest segmentation, and a fully open-source architecture for reproducibility. The system accepts DICOM, NIfTI, and ZIP inputs, and includes optional YAML-configured biomarker simulations (TriOx) and conceptual clinical risk-factor modules. These components are exploratory and intended as proof-of-concept simulations rather than validated clinical predictors. An automatic dual-audience reporting component generates structured summaries for both clinicians and patients. Internal evaluations used 10 000 Monte Carlo simulations, incorporating literature-derived Helicobacter pylori prevalence estimates and imaging statistics from the TCGA-STAD dataset. To explore potential deployment variability, experiments were conducted under simulated "urban" (higher-quality imaging, complete metadata) and "rural" (lower resolution, partial metadata) screening conditions. For external assessment, we applied the pipeline to 773 independent CT scans from the AbdomenCT-1K TumorSubset. Because this dataset lacks segmentation ground truth, the experiment was used to evaluate cross-dataset reproducibility and stability, without retraining or parameter tuning, thus reflecting reproducibility rather than accuracy assessment. An anonymized English summary of the external validation process is provided in Supplementary File S1. All source code, configuration files, and example data are publicly available to support end-to-end transparency and reproducibility.

Across 10 000 Monte Carlo simulations representing urban and rural screening conditions, Banana 0.9 produced a simulation-derived mean AUC of 0.87 (95% CI 0.84-0.90). Estimated computational demand was reduced by more than 80%, with model-based projections suggesting a ~ 60% decrease in average per-patient screening costs relative to conventional GPU-dependent workflows, an estimate based on assumptions regarding typical hardware pricing, device lifespan, and energy consumption. Simulated detection rates increased from 70% to 85% under "urban" conditions and from 65% to 80% under "rural" conditions. For external assessment, Banana 0.9 processed 773 independent CT scans from the AbdomenCT-1K TumorSubset with 100% successful execution and without retraining or parameter adjustment. Although this dataset does not provide segmentation ground truth, no instability or failure modes were observed relative to internal simulations, indicating reproducible operation across heterogeneous imaging domains.

Banana 0.9 offers an open, transparent, and GPU-free imaging framework aimed at improving reproducibility and accessibility in gastric cancer screening workflows. Using internal Monte Carlo simulations and external execution on an independent CT dataset, the system demonstrated consistent and reproducible operation without retraining or parameter adjustment, providing preliminary evidence of stability across heterogeneous imaging conditions. While the present evaluation relies on simulated performance estimates and non-annotated external data, the modular architecture, openly available codebase, and low computational requirements position Banana 0.9 as a practical starting point for future extensions toward clinically validated, multi-cancer CT screening tools aligned with FAIR data principles and global health needs.