Interpretable vision transformer with dual loss and Grad-CAM for iris-based early lung damage detection
Abstract
Iridology infers organ health from the appearance of the iris; its clinical record is contested, but iris photography combined with learned representations allows the underlying question-whether disease-correlated structure exists in iris images-to be tested against external clinical labels. We present an interpretable framework combining a vision transformer (ViT) trained with a dual cross-entropy and triplet objective with a quantitative evaluation of its gradient-weighted class activation mapping (Grad-CAM) explanations, for binary classification of lung damage from iris images labelled by independent clinical diagnosis. On a held-out set of 356 single-participant images the model reaches 0.9249 precision and 0.9242 recall, matching a ResNet-50 baseline and a single-loss ViT to within two images at 21% lower datacentre latency. The principal contribution is the measurement of interpretability rather than its assertion: true-positive saliency maps are causally faithful, reproducible and class-discriminative, while their 0.9315 overlap with false-positive maps bounds what they may be taken to mean - they show where the model finds positive evidence, not where pathology is. We additionally audit acquisition confounding and find that image-quality statistics alone separate the classes at area under the receiver operating characteristic curve (AUROC), establishing a floor that iris-based screening studies should report and that, to our knowledge.
Keywords
Gradient-weighted class activation mapping; Iridology; Lung damage; Pre-screening; Vision transformer
Full Text:
PDFDOI: https://doi.org/10.11591/eei.v15i6.13484
Refbacks
- There are currently no refbacks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Bulletin of Electrical Engineering and Informatics (BEEI)
ISSN: 2089-3191
,
e-ISSN: 2302-9285
This journal is published by the
Institute of Advanced Engineering and Science (IAES)
.