Attention modules add nothing on nine-class chest X-rays: five add-ons tie an identical ResNet-50 across ten seeds

With the same ResNet-50 backbone, data split, optimizer and schedule, and only the inserted module changed, none of the five lightweight attention or graph modules raised macro-AUC on nine-class chest radiograph classification. The standard recipe adapts an ImageNet-pretrained ResNet backbone and then inserts channel attention, spatial-channel attention or graph layers, reporting accuracy gains that are rarely subjected to paired statistical testing. The baseline reached 0.7990±0.0023 per-seed macro-AUC, against SE 0.7991, BAM 0.7999, GCN 0.7977, CBAM 0.7962 and GAT 0.7936, with all six ensemble confidence intervals overlapping; CBAM (−0.0028, p=0.0195) and GAT (−0.0053, p=0.0020) were significantly lower after Benjamini–Hochberg correction, but the 95% confidence interval of every paired difference lay inside ±0.01 AUC.

Where the module is inserted changes the outcome: the original in-branch placement neither helps nor harms, while gating each block's complete output degrades performance. The main benchmark inserted each module once after the backbone, so the original per-block placement needed a separate test. SE and CBAM inside the residual branch of every trainable block reached 0.8003±0.0025 and 0.8006±0.0026, that is +0.0013 and +0.0016 against the baseline and below each comparison's minimum detectable effect; gating each block's complete output dropped them to 0.7716±0.0082 and 0.7286±0.0184, that is −0.0274 and −0.0703, negative in 10 of 10 seeds (p=0.0020).

The absence of gain does not depend on training-set size, backbone choice or the patient overlap of the image-level split. The main benchmark used 9,532 training images and a single ResNet-50 backbone, leaving small-data, backbone-specific and leakage explanations open. Raising the training set to 22,898 images lifted the baseline to 0.8112±0.0019 with no module gaining; across ResNet-18, EfficientNet-B0 and MobileNetV2, 19 of 21 module-versus-baseline comparisons fell inside the 0.01 margin, 2 were worse and none was better; a patient-level split lowered the baseline to 0.7893±0.0033 and left the conclusions unchanged.

The conclusions cover one dataset, NIH ChestX-ray14, at 224×224 input, in a partially frozen fine-tuning regime that freezes the first two residual stages, on a nine-class single-label chest radiograph task. The larger training set, three additional backbones and the patient-level re-split all support the absence of gain, but the non-ResNet backbones had the first half of their feature hierarchy frozen proportionally rather than matched on parameter count, and EfficientNet-B0 already contains squeeze-and-excitation inside its MBConv blocks, so its baseline is not a module-free control. The per-block output-gating penalty was pronounced only on ResNet-50 and is not generalized to the other backbones.

The validation set served both early stopping and reporting, so all six models carry a mild optimistic bias and the absolute macro-AUC values should not be read as held-out performance. The main benchmark uses an image-level split in which 1,009 patients contribute images to both sets, covering 1,350 validation images (56.7%), and this overlap attenuates between-model differences in the same direction as the equivalence conclusion. Pneumonia and cardiomegaly have few images, so their class-specific estimates are unstable across seeds. Each module was implemented in one reasonable way, without per-module tuning of learning rate, graph projection placement or graph layer count. Whether the task ceiling is set by ChestX-ray14 label noise rather than model capacity would require re-annotated labels to distinguish.