Energy-preserving Quantum-inspired Wavelet Features for Mri Abnormality Detection

Abstract Identification of abnormalities correctly in magnetic resonance imaging (MRI) is a challenging problem, especially in clinical situations with limited annotated data. Here, a quantum-inspired multi-scale wavelet energy framework is proposed for automated MRI abnormality detection and localization. Framework combines principles from quantum mechanics, with classical signal processing, to achieve stable, energy-preserving and interpretable representations of features. Firstly, MRI images are pre-processed and decomposed in overlapping patches using a sliding-window strategy. Each patch is projected onto a state representation in a quantum-inspired way, by constraining the energy to 1, which allows for the analysis in a Hilbert space. Orthonormal wavelet transforms (Haar and Daubechies-2 (db2)) are used as energy preserving unitary transformation to project the normalized patches to frequency subspaces. High frequency wavelet sub-band energies are computed for characterization of structural irregularities linked to pathological tissue. To overcome the problem of the size variability of the damaged tissue, a multi-scale analysis is conducted based on 8X8, 16X16 and 32X32 pixel patch sizes. Spatial energy maps are created and fused for use in localization of damaged tissue. A relatively small and understandable feature vector is created through the fusion of patch-level statistics, wavelet energy distributions, high-frequency to low-frequency energy ratios, and damaged tissue concentration measures. Classification is done with support vector machine (SVM) and radial basis function kernel. Experimental results show a definite scale-wavelet dependency. Extremely small patches (p = 8) result in high sensitivity, but less specificity from noise amplification. The db2 wavelet gives the best performance at p = 16, where it achieves an accuracy of 72.60%, a sensitivity of 89.13%, and an F1-score of 0.8039, which shows that the capability of detecting the abnormality is very good. In contrast, the Haar wavelet shows more stable behaviour at larger patch sizes (p = 32) which offers a balanced sensitivity/specificity trade-off. The proposed framework results highlights the importance of adaptive scale and wavelet selection in energy-based MRI abnormality detection.