What Matters for Lightweight Multi-Label Chest X-Ray Classification? A Controlled Multi-Seed Study of Attention, Loss, Width and Resolution

Abstract

Attention modules and imbalance-aware losses are often credited with gains in multi-label chest X-ray classification, but their effects are seldom separated from training length, model width, input resolution and seed-to-seed variation. We integrate three existing components, squeeze-and-excitation (SE) channel attention, the class-specific residual attention (CSRA) classifier and the asymmetric loss (ASL), into a lightweight residual network (SECA-Net) and evaluate them on the 14-label ChestMNIST benchmark in 87 training runs. A 2×2×2 factorial ablation with five seeds, a loss study, a width study, three lightweight reference models and a 64- and 128-pixel resolution study were all trained to convergence under one protocol. None of the three components changed macro-AUROC significantly (seed-blocked three-way ANOVA: SE p = 0.84, CSRA p = 0.29, ASL p = 0.86), and their combination did not significantly beat the baseline at width 64 (+0.0018, p = 0.085) or at 64 and 128 pixels. In contrast, input resolution raised macro-AUROC from 0.789 at 28 pixels to 0.815 at 128 pixels with the same 0.70 M-parameter network, training to convergence instead of eight epochs added 0.014, and widening the network from 16 to 64 base channels added 0.007. ASL and weighted cross-entropy left discrimination unchanged but made probabilities poorly calibrated (class-wise expected calibration error 0.31 versus 0.006 for cross-entropy), which per-class Platt scaling corrected. Differences of up to 0.01 AUROC between single runs can arise from seed-to-seed variation and GPU non-determinism alone, so multi-seed evaluation is essential. Code, trained weights and machine-readable results are released.