TB Detection System

01 / 09

02 / About

About
the project.

An explainable deep-learning prototype that screens frontal chest X-rays for TB-consistent abnormalities.

800Real chest X-rays

DenseNet-121Transfer-learned backbone

0.892Holdout ROC-AUC

2Independent hospital sources

The system

What it does.

Binary screeningNormal vs TB-consistent abnormality
Image-onlyThe X-ray decides — form answers never shift the score
ExplainableGrad-CAM overlay shown beside every result
Calibrated outputProbability plus a documented operating threshold
Quality-awareWarns when an image is unfit for reliable screening
Privacy-firstProcessed in memory; nothing stored by default

How it works

The pipeline.

01IngestPNG / JPG / DICOM, validated on upload
02NormaliseGrayscale, aspect-preserving pad, 224 × 224
03EnhanceCLAHE & lung-ROI tested as ablations, not assumed
04ClassifyDenseNet-121, ImageNet weights, sigmoid head
05ThresholdScreening flag from a pre-declared cut-off
06ExplainGrad-CAM heat map over the original X-ray

The ML model

Under the hood.

ArchitectureDenseNet-121 · ImageNet transfer learning
HeadSingle sigmoid output, BCE with positive weighting
Input224 × 224, grayscale → 3-channel, ImageNet-normalised
OptimiserAdamW · lr 2e-5 · weight decay 1e-4 · seed 42
DataShenzhen 662 + Montgomery 138 · 652 train / 148 held out
AugmentationMild affine & brightness — training split only
BaselinesResNet-50 and EfficientNet-B0 for comparison

Held-out results · threshold 0.50

Set	Sens.	Spec.	AUC
Shenzhen n=120	70.5%	94.9%	0.892
Montgomery n=28	41.7%	100%	0.828

Montgomery is a fully external source. The sensitivity drop is reported, not smoothed over — it is exactly why this stays a screening prototype.

Honest limits

What it cannot claim.

Not a diagnosis — screening support only
Two hospital sources; multi-site validation pending
Operating threshold not yet fixed on a validation split
Grad-CAM explains attention, not clinical correctness

The team

Meet the team.

Dept. of Computer Science & Engineering · Guide: Prof. Vikranth B.M

01 / Developer

Nihaan Jonas

1BM23CS213 · Sec 7D

02 / Data & ML

Lakshith P

1BM23CS164

03 / Report & design

Kishen Deepak

1BM23CS152 · Sec 7C

04 / Testing & research

K C Sai Nithin

1BM23CS130 · Sec 7C

Research demonstration for academic assessment. Not a medical device and not for clinical or treatment decisions.