An explainable frequency-domain ML pipeline for lung histopathology classification

Background Accurate subtype classification of lung histopathology is clinically valuable, but many high-performing deep models are opaque and compute-intensive. We target a transparent, CPU-feasible pipeline that maintains competitive accuracy on a standard benchmark. Method Red, Green, Blue (RGB) patches are resized to 384 × 384, converted to Hue, Saturation, Value (HSV), and a 2D Fast Fourier Transform (FFT) is applied per channel. From the conjugate-symmetric magnitude, we extract 41 descriptors per channel (statistics, band-powers/ratios, entropy/energy, percentiles, edge statistics, centroid, Gini, radial means), yielding 123 Color Frequency Spatial Descriptor (CFSD) features. Mutual Information ranking with a data-driven cutoff reduces this to 54 features. Eight classifiers (XGBoost, LightGBM, CatBoost, Random Forest (RF), Extra Trees (ET), Gradient Boosting (GB), Bagging, Decision Tree (DT)) are trained on an 80/20 split of the lung subset of LC25000 using 5 non-overlapping folds. Metrics include Accuracy, Precision, Recall, F1-score, Jaccard, MCC, Cohen’s κ, Specificity, and area under the curve (AUC). Results With 123 features, eXtreme Gradient Boosting (XGBoost) achieves 97.85 ± 0.25% accuracy; using the 54-feature subset yields 97.26 ± 0.25% with <1% F1-score decrease. Mean feature-extraction time drops from ~72 ± 4 ms to ~51 ± 3 ms per image (≈28% faster), with end-to-end per-image latency reductions of ~21 ms across models. Confusion exists between adenocarcinoma and squamous cell carcinoma; normal tissue is most separable. Findings are showing robust results against state-of-the-art methods.