Lightweight Spatio-Sequential Feature Modeling for Parkinsonian Pattern Recognition in Functional MRI

Parkinson's disease (PD) is a progressive neurodegenerative disorder, and Functional Magnetic Resonance Imaging (fMRI) can provide complementary information for pattern analysis. This study evaluates a compact spatio-sequential framework for recognizing PD-related patterns in a public set of 279 Digital Imaging and Communications in Medicine (DICOM) fMRI images from healthy controls and patients with PD. Preprocessing included intensity whitening, background stabilization, recovery of 49 slice channels, removal of one low-information terminal slice, packing of the remaining 48 slices into three channels, and resizing to 224 × 224 pixels. Training used horizontal reflection and small in-plane rotations. ImageNet-pretrained MobileNetV2 produced 2,000 descriptors; Linear Discriminant Analysis (LDA)-based ranking retained 1,000, which were classified by two Bidirectional Long Short-Term Memory (BiLSTM) layers with 128 hidden units and 0.4 dropout. Across ten runs, mean test accuracy was 98.3670%, precision 98.3698%, recall 97.3529%, F1-score 97.8507%, and Receiver Operating Characteristic Area Under the Curve (ROC-AUC) 0.9976. A representative split correctly classified 66 of 67 images. The results indicate strong internal discrimination, although image-level splitting and limited external validation restrict clinical interpretation.