<p>Accurate identification of lung infections and abnormalities from medical images is essential for early diagnosis and effective clinical decision-making. However, reliable detection remains challenging due to heterogeneous visual characteristics across imaging modalities, particularly Positron Emission Tomography (PET), and Computed Tomography (CT). Inconsistent representation of anatomical structures and texture patterns across modalities often results in tissue misclassification and imprecise localization of early-stage tumors or solid masses such as atelectasis. To address these limitations, this paper proposes TexFNet (Texture Fusion Network), a novel multimodal learning framework that performs texture-driven spatial anatomical variation modeling for precise lung disease localization. TexFNet operates through three sequential stages: image acquisition and preprocessing, spatial anatomical variation analysis, and texture-fused image generation. A YOLO-inspired bilayer learning architecture (TexNet) is introduced to explicitly distinguish variation and non-variation texture patterns between PET and CT images using cosine-similarity-guided feature learning. This design enables region-aware fusion by suppressing irrelevant anatomical variations while enhancing pathological texture discrepancies. Unlike conventional feature-level or decision-level fusion approaches, TexFNet performs fusion at the texture–spatial level, allowing fine-grained localization of diseased lung regions. The proposed framework is evaluated on publicly available PET–CT datasets using accuracy, precision, recall, and mean average precision (mAP) metrics. Experimental results demonstrate that TexFNet achieves improvements of 12.51% in accuracy, 12.38% in precision, and consistently higher mAP compared to state-of-the-art multimodal fusion and deep learning models. These results confirm the effectiveness of texture-driven spatial fusion for robust and early lung disease detection.</p>