A leakage-controlled multiscale contextual involution-inception framework for spectral-spatial classification of...

Early and reliable assessment of neonatal conditions is important in neonatal intensive care units, where non-invasive imaging approaches may provide complementary information for clinical evaluation. Hyperspectral imaging provides a non-invasive means of capturing spatially resolved spectral information from neonatal tissue; however, its high dimensionality and the limited availability of clinical samples make robust model development challenging. This study proposes the Multiscale Contextual Involution-Inception Network (MC-I²Net), a leakage-controlled framework for neonatal HSI classification. The model combines spatially adaptive involution operators, Inception-based multiscale feature extraction, and contextual feature fusion to learn compact and discriminative spectral-spatial representations. A balanced six-class neonatal hyperspectral imaging dataset containing 240 samples was evaluated, including coarctation of the aorta, intracranial hemorrhage, necrotizing enterocolitis, pneumothorax, respiratory distress syndrome, and healthy control groups. Fixed signal-conditioning operations were applied using predefined processing rules, whereas PCA was treated as a data-adaptive transformation. PCA was fitted exclusively on each training fold, and the learned transformation was subsequently applied to the corresponding held-out test fold. All models were assessed using the same 5-fold cross-validation protocol repeated 10 times. MC-I²Net achieved an overall accuracy of 94.38 ± 4.60% and a Macro-F1 score of 94.35 ± 4.57%, with only small numerical differences in OA relative to MC-3DCNN, ViT-HSI, and Swin-HSI. The model used fewer trainable parameters than MC-3DCNN, ViT-HSI, and Swin-HSI and substantially fewer FLOPs than MC-3DCNN, supporting a favorable performance-complexity profile rather than a large predictive-performance margin. An additional anatomically controlled analysis restricted to PTX and RDS samples acquired from the same thoracic region retained an OA of 95.75 ± 6.49% and a Macro-F1 score of 95.66 ± 6.68%. Sequential nested cross-validation, with PCA dimensionality and involution hyperparameters selected using training data only, yielded a pooled outer out-of-fold OA of 95.00% and a Macro-F1 score of 94.98%. Ablation, sensitivity, calibration, nested-validation, runtime, and saliency analyses provided complementary evidence regarding architectural behavior, robustness, computational characteristics, and preliminary explainability. Overall, these findings support MC-I²Net as a proof-of-concept framework for neonatal HSI classification and provide a methodological basis for future prospective, multicenter, and cross-device evaluation.