ChestInsight — MATLAB-only chest X-ray project

This is a MATLAB-only academic prototype for multi-label analysis of frontal chest X-rays. It evaluates ten findings independently:

Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, and Edema.

The project does not need Python, PyTorch, Streamlit, a web browser, or a GPU. It runs in MATLAB R2026a using base MATLAB functions. The code does not need the Deep Learning Toolbox or Image Processing Toolbox.

Run the application

1. Double-click Start MATLAB App.cmd.
2. Click Use built-in example or 1. Choose X-ray.
3. Select the finding you want to inspect.
4. Click 2. Analyze in MATLAB.
5. Read the ten scores and view the MATLAB contrast-enhanced image and saliency map. A PNG, MAT file, JSON, CSV, and images are automatically saved under reports/matlab_only_runs.

Use de-identified frontal AP/PA PNG or JPEG chest X-rays only.

Retrain the MATLAB model

Double-click Train MATLAB Model.cmd.

It builds MATLAB features from data/images, trains a two-hidden-layer ReLU neural network, selects each finding threshold using validation patients, and evaluates once on untouched test patients. The resulting model is saved at:

models/chestinsight_matlab_model.mat

The original NIH training-image collection is not stored in this repository. Download it from the source linked in data/SOURCE.md, then place the images in data/images before retraining. The included saved model, demo X-rays, manifest, and extracted MATLAB features are enough to run the application.

The first run creates data/matlab_features.mat; later runs reuse it. Delete that file only when the dataset images or manifest are changed.

To train from the MATLAB Command Window:

cd('C:\Users\Mathews\Documents\ChatGPT\BME')
addpath('matlab')
chestinsight_train('MaxEpochs',10)
chestinsight_app

How it works

1. MATLAB converts the X-ray to grayscale, normalizes it, and resizes it to 32×32 pixels.
2. A custom two-hidden-layer neural network produces ten independent sigmoid scores. Several findings can be above threshold at the same time.
3. MATLAB calculates an input-gradient saliency map for the selected finding. It shows which image pixels most affect that output; it is not a lesion boundary.
4. MATLAB creates a display-only contrast-enhanced view, selected-finding result panel, score chart, PNG, CSV, MAT file, and JSON summary.

Measured MATLAB-only pilot result

The included model was trained on 4,312 images from 1,306 patients, selected with 1,096 images from 327 validation patients, and evaluated on 2,560 images from 162 untouched test patients.

Metric	Result
Macro AUROC	0.593
Macro average precision	0.117
Macro F1	0.168

See reports/matlab_only/test_metrics.csv for results for every finding.

Limits

This is an academic research prototype. It uses a small, non-random pilot subset of NIH ChestX-ray14 and report-derived labels. Scores are not calibrated clinical probabilities, and the system is not validated for diagnosis or treatment.