Improving Accuracy and Evaluating the Generalizability of Medical Image Classification Using Pseudo-Invisible-Light Images Generated from RGB Images

We extend our previous study on extended 5-channel images, which add two pseudo-invisible-light channels, estimated by spectral reconstruction (MST++) and polynomial extrapolation, to the RGB channels. We first ask whether the accuracy gain reflects prior knowledge contributed by the externally trained spectral-reconstruction model or effects that any 5-channel input derived from RGB would produce. For skin cancer classification with a CNN trained from scratch, the proposed input is compared with grayscale-based extensions, duplicated RGB channels, learned pointwise transforms of RGB, and learned projections of the reconstructed spectrum. Duplicating RGB channels alone raised accuracy from 86.49% to 88.27%; learned transforms of RGB did not significantly exceed this level, whereas the proposed input (89.94%) exceeded both; learned projections of the reconstructed spectrum did not differ significantly from it. We then re-evaluate the method on diabetic retinopathy, gastrointestinal, and glaucoma classification with ImageNet-pretrained ResNet50, with wavelengths selected on validation data. The proposed input exceeded RGB by 1.5–2.1 percentage points in all three tasks, as did most wavelength combinations, but these gains were not significant and not distinguishable from the RGB-duplication control; an effect specific to the proposed input remains unestablished there. Computational cost is also reported.