Specimen-Wise Evaluation of Acute Leukemia Subtyping Using Microscopic Hyperspectral Imaging.
文献摘要
Background Microscopic hyperspectral imaging provides label-free, wavelength-resolved optical signatures that may support hematologic smear screening. We evaluated whether a compact spectral model could discriminate acute lymphoblastic leukemia, acute myeloid leukemia, and healthy controls while minimizing information leakage. Methods Hyperspectral microscopy acquired 128 bands over the nominal 370-980 nm detector range. After dark/white reference calibration, edge bands with poor signal-to-noise were excluded and downstream analyses were restricted to the usable 400-900 nm interval; no min-max normalization or SNV was applied after reflectance calibration. A compact one-dimensional convolutional neural network was evaluated on 48 bone-marrow smear specimens (acute lymphoblastic leukemia, n=18; acute myeloid leukemia, n=15; healthy controls, n=15) using specimen-wise stratified five-fold cross-validation. Pixel-level probabilities were aggregated to specimen-level predictions. Results After restricting the analysis to 400-900 nm, the model achieved an accuracy of 0.921±0.011, macro-F1 of 0.910±0.010, and macro-area under the receiver operating characteristic curve of 0.962±0.007. Principal component analysis showed class-related structure in the retained spectral window, and cross-validated PCA-LDA provided a lower-complexity comparator for sensitivity, specificity, and accuracy. Conclusion Quality-controlled visible/near-infrared microscopic hyperspectral signatures supported three-class discrimination under leakage-controlled specimen-wise validation. The analysis explicitly removes low-signal edge bands and provides PCA/loading and discriminant-analysis baselines in addition to the compact CNN.