Iridescent screens frontal chest X-rays with DenseNet-121 models and shows where each result comes from, using Grad-CAM, LIME and SHAP, with plain-language reliability checks. Everything runs in your browser: the image never leaves your device.
Runs on your device. No upload, works offline.
The app reads the image exactly as the models were trained to, runs the models on your device, and turns their output and explanation maps into a short, checkable summary.
Grayscale, resized the same way as in training. PNG or JPEG, any size.
Two DenseNet-121 models: a 14-condition screen for adults, and paediatric pneumonia. The original and a mirrored copy are averaged.
Grad-CAM shows where the model looked; LIME and SHAP test which image regions change the result.
Is the attention inside the chest? On film edges or text markers? Do the methods agree? How close is the result to the threshold?
All results are on data the models never saw during training, split by patient so that no person appears in both. Pneumonia metrics are the mean ± standard deviation of three independent training runs.
| Model | AUROC | Sensitivity | Specificity |
|---|---|---|---|
| Initialised from our NIH ChestX-ray14 pretraining used in the app | 0.997 ± 0.001 | 98.3% | 97.1% |
| Initialised from ImageNet | 0.996 ± 0.001 | 97.4% | 98.0% |
| ImageNet, lung-cropped images | 0.988 ± 0.001 | 96.4% | 94.1% |
The two full-image models are statistically equivalent in accuracy; their difference shows in where they look (below).
DenseNet-121 trained on 78,186 adult X-rays from NIH ChestX-ray14 and tested on the official patient-wise test list of 25,596 images. Mean AUROC 0.81, from 0.69 (infiltration) to 0.92 (emphysema). Scores are corrected for the weighted training loss.
On 984 lesions outlined by radiologists, the hottest point of the Grad-CAM map fell inside the box 34% of the time: 4.5 times chance. It works well for large findings such as an enlarged heart (89%) and poorly for small nodules (1%).
Accuracy alone would have hidden these. They come from measuring the explanation maps on 100 test images per model.
For pneumonia predictions, the ImageNet model put 8.0% of its Grad-CAM attention in the two top corners, where view markers such as "AP" are printed, against 4.9% for normal predictions. Cropping to the lungs removed this but lowered specificity from 98.0% to 94.1%.
Starting from our own NIH ChestX-ray14 model instead of ImageNet cut corner attention for pneumonia predictions to 4.8% while keeping accuracy. The benefit of domain pretraining here is in what the model attends to.
When the top of the network is re-initialised, Grad-CAM maps change completely, but SHAP maps stay about 60% the same: they partly reflect image structure rather than what the model learned. The three methods also often disagree, so the app never presents one map as "the" reason.
Iridescent is a research prototype built as a final-year project. It is not a medical device and does not give a diagnosis.
The models are downloaded once and run inside the browser with ONNX Runtime Web, on the GPU when it passes an accuracy check and otherwise on the CPU. The X-ray is never uploaded, results are stored only on the device, and the app works offline after the first use.
Open the appExplainable AI for Medical Image Diagnosis: a final-year B.E. project in Electronics and Telecommunication Engineering. Data: Kermany et al. (2018), Chest X-Ray Images (Pneumonia), CC BY 4.0; Wang et al. (2017), NIH ChestX-ray14. Lung segmentation in the evaluation: TorchXRayVision.