Small cell lung cancer patches get a spectral-bilinear transformer: 0.9697 accuracy at a third of MLP-Mixer's parameters
Synopsis
Classifying small cell lung cancer histopathology requires resolving high-frequency nuclear texture and long-range tissue organisation at once, which fixed-kernel convolutions and pure transformers each handle poorly. SCLC-SBT puts frequency-domain channel recalibration in its first three stages and a dual convolutional and 2D selective state-space branch in its deepest stage, reaching 0.9697 accuracy, 0.9728 F1 and 0.9971 AUC on a 1,974-patch binary benchmark. Against the six strongest baselines the difference is not statistically significant on the 297-sample test set, so its clear advantage is competitive accuracy at about 17M parameters and 7.0G FLOPs.
Interpretation
Small cell lung cancer patch classification no longer relies on spatial convolution or global self-attention alone, handling frequency-domain channel statistics and long-range spatial dependencies inside one hierarchical network. Convolutional networks were constrained by fixed kernel sizes and struggled with tissue-level architectural disorganisation, while pure transformers scale quadratically with sequence length and miss fine nuclear texture. On a patient-level split of 1,974 H&E patches with a 297-sample test set, SCLC-SBT reached 0.9697 accuracy, 0.9728 F1 and 0.9971 AUC, with the best point estimate on accuracy, recall, F1, AUC and MCC among eleven compared architectures including DeiT, MLP-Mixer, DaViT and CAFormer.
The first three stages use re-parameterizable hybrid convolution blocks for multi-scale local features, with frequency-domain channel attention inserted among them. The spectral-channel attention averages the 2D fast Fourier transform magnitude of each channel over spatial dimensions into an energy descriptor, then a two-layer MLP turns it into channel weights. Only half the channels undergo spectral recalibration while the rest pass through unchanged; replacing it with standard channel attention lowered single-run accuracy by 0.0135 and MCC from 0.9386 to 0.9113.
The deepest stage runs progressively dilated convolution in parallel with a 2D selective state-space scan, fused by multiplicative gating from a depthwise bilinear attention module. The convolutional branch expands its receptive field with dilation rates d, 2d and 3d, while the linear branch scans the feature map along four spatial directions and averages them to model global context at linear complexity. On the 297-sample test set, removing the DBCLIT convolutional branch cut accuracy by 0.0269, the largest single-module drop; across five random seeds, removing that branch or the SS2D branch remained statistically significant against the full model.
Perspective
Results come from 1,974 H&E patches extracted from 104 whole-slide images of 89 patients at a single institution, split by patient into a 297-sample test set, so they reflect between-patient generalisation rather than transfer across institutions, scanners or staining protocols. Throughput on one RTX 4060 is 417 img/s at batch size 8 and falls to 99 img/s for single-sample inference.
The authors state that against the six strongest baselines, DeiT, MLP-Mixer, CAFormer, Swin-T and the tied PoolFormer and DaViT pair, the difference is not statistically significant on the 297-sample test set, so the accuracy ranking is competitive rather than conclusively superior. Multi-seed repetition covers only the five component-removal configurations, design-choice comparisons remain single runs, and generalisation across institutions, scanners and staining protocols awaits multi-institutional study.