We introduce FuzzyCoMobile, a clinically oriented, ROI-aware framework for multi-label chest X-ray (CXR) classification. The method ensembles three complementary backbones MobileViT-S, ConvNeXt-Tiny, and CoAt-Lite-Medium via a class-wise optimized fuzzy aggregation that adaptively weights model outputs per label. To reduce background confounders and focus on plausible lesions, FuzzyCoMobile integrates YOLOv12-based lesion detection with SAM-refined ROI cropping trained on VinBigData, while classification is performed on NIH ChestXray14. On the 7 overlapped labels (patient-wise 70/10/20 split; \(224\times 224\) input), the ensemble consistently surpasses the strongest single model across four matched scenarios, achieving macro ROC–AUC of 0.847 and multi-label accuracy of 95.6%. Beyond discrimination, we quantify explanation quality: ensemble heatmaps attain a Pointing-Game score of 0.85. Explanations are generated using Grad-CAM, Score-CAM, and LIME, providing interpretable, ROI-aligned rationales and a practical pathway toward trustworthy CXR decision support.