SpecAtt-Net: Time-Frequency-Based Image Representations of Slice-Level Radiomic Features for Explainable Lung Cancer...

Background: Extracting radiomic features from volumetric imaging data for lung cancer classification is limited by high dimensionality and the black-box nature of deep learning models. Traditional methods may overlook dependencies between slices, leading to the loss of sequential spatial information. Even when classification is correct, the model's explanatory power may remain insufficient. Objective: This study proposes SpecAtt-Net, which processes spectral representations of slice-level radiomic data from computed tomography (CT) scans of lung cancer patients. A slice-based explainability approach is also presented to identify the slice groups used in model decision-making. Methods: Slice-level radiomic data were preprocessed, and slice-position-frequency representations were generated using the Short-Time Fourier Transform (STFT). This enables the model to capture spectral textures and spatial transitions from 720 × 288 feature maps. SpecAtt-Net employs a dual-attention mechanism to focus on distinctive radiomic features. To address the lack of slice-level labels, Slice-Wise Activation Mapping (SWAM), a post hoc interpretability technique, was developed. SWAM converts two-dimensional (2D) activation maps into one-dimensional (1D) importance vectors and provides a weakly supervised indication of the slice groups most relevant to the model's decision. Results: SpecAtt-Net achieved competitive performance in accuracy, precision, recall, and F1-score compared with the evaluated reference architectures. SWAM highlighted slice-sequence regions contributing strongly to model predictions, providing preliminary evidence for interpretability. Conclusions: SpecAtt-Net provides a compact and interpretable approach for lung cancer subtype classification using attention-based feature modeling. SWAM offers an exploratory visualization of decision-relevant sequence regions and may support future radiological interpretation and validation studies.