Artificial intelligence (AI) is transforming oncologic imaging by enabling automated, accurate, and scalable diagnostic decision support. However, successful clinical translation requires not only strong predictive performance but also interpretability, transparency, and clinician confidence. We present an attention-guided deep learning framework for automated slice-level classification of parathyroid adenoma (PTA) from arterial-phase computed tomography (CT) images. The framework incorporates lightweight hierarchical attention mechanisms within residual neural networks to enhance feature representation and contextual understanding while maintaining computational efficiency for real-world deployment. Three novel architectures were developed: the Residual Block Light Attention Network (Res-BLANet), Residual Stage Light Attention Network (Res-SLANet), and Residual Layer Light Attention Network (Res-LLANet). Models were trained and evaluated on a rigorously curated, expert-annotated dataset of 63 patients, including 39 pathologically confirmed PTA cases, with 350–450 slices per patient. Training employed Hounsfield unit normalization, extensive data augmentation, and 10-fold cross-validation to ensure robust performance assessment. A key innovation is the integration of hierarchical attention modules that generate attention maps at multiple network levels, enabling qualitative visualization of diagnostically relevant regions and providing exploratory insight into the model’s decision-making process. Experimental evaluation showed strong diagnostic performance. Res-SLANet achieved the highest overall accuracy (87.37%), precision (92.56%), and F1-score (86.69%), while Res-LLANet attained the highest sensitivity (96.0%) on independent testing. These results should be interpreted as preliminary, given the limited size of the independent patient-level test cohort (n = 3). Res-BLANet delivered substantially faster inference with minimal computational overhead. These findings demonstrate that hierarchical attention-guided deep learning can achieve accurate and computationally efficient PTA detection from CT imaging while providing qualitative visual insights into the model’s decision-making process. The proposed framework represents a promising approach toward more interpretable AI-assisted diagnostic systems for oncologic imaging.

Reliable four-class decoding remains challenging for functional near-infrared spectroscopy (fNIRS) brain–computer interfaces (BCIs), because a single feature domain captures only part of the hemodynamic response. This study proposes an Adaptive Multi-Domain Feature Fusion (AMFF) framework that combines six temporal hemodynamic descriptors with ten fNIRS-adapted Mel-frequency cepstral coefficients (MFCCs) and learns sample-dependent weights for the two domains before classification with a multilayer perceptron. The framework was evaluated on an open-access dataset of 29 participants performing left- and right-hand motor imagery, mental arithmetic, and rest, using subject-independent leave-one-subject-out cross-validation. AMFF achieves a mean accuracy of 95.6%, with precision, recall, and F1-score of 94.6%, 95.9%, and 95.3%, respectively, outperforming temporal only (90.7%), cepstral-only (95.0%), and direct concatenation (95.2%) representations, with a modest but consistent gain over direct concatenation across participants. Friedman and Holm-corrected Wilcoxon signed-rank tests confirmed significant differences for all four metrics; AMFF exceeded static concatenation for every participant, although by a small margin (0.2–0.7 percentage points). Most of the gain over temporal descriptors therefore comes from the cepstral domain, with adaptive weighting adding a small but consistent refinement. MFCC parameter-sensitivity, coefficient-ablation, and descriptor-ablation analyses further characterize the contribution of each domain.