LungLens

Inspiration

Pneumonia is one of the leading causes of death in young children worldwide, and it is highly treatable when caught early. The bottleneck is often not treatment but diagnosis: in many clinics there are too few radiologists, and a chest X-ray can wait a long time to be read.

AI can help, but most medical AI tools are black boxes. They output "pneumonia: 94%" with no explanation, and a doctor has little reason to trust a number they can't question. I wanted to build something that doesn't just predict, but shows its reasoning and admits when it isn't sure. That became LungLens: see what the AI sees.

What it does

Upload a chest X-ray and LungLens returns:

• A prediction (Normal vs. Pneumonia) with a confidence score
• A Grad-CAM heatmap highlighting the lung regions that drove the decision
• An "Uncertain, please consult a doctor" result when confidence is too low, instead of forcing a guess

It is designed as a triage aid to help prioritize scans, not to replace clinicians.

How I built it

• Model: transfer learning with a pretrained [ResNet18 / MobileNetV2] in PyTorch, fine-tuned on the public Chest X-Ray Images (Pneumonia) dataset
• Data handling: the provided validation set was too small to be reliable, so I created my own stratified train/validation/test split, with augmentation (rotation, brightness, flips) to improve generalization
• Class imbalance: pneumonia images outnumber normal ones, so I used a weighted loss:

$$\mathcal{L} = -\frac{1}{N}\sum_{i=1}^{N} w_{y_i}\,\log p_{\theta}(y_i \mid x_i)$$

• Explainability: Grad-CAM weights each feature map $A^k$ by the average gradient of the class score $y^c$:

$$\alpha_k^c = \frac{1}{Z}\sum_{i}\sum_{j}\frac{\partial y^c}{\partial A^k_{ij}}, \qquad L^c_{\text{Grad-CAM}} = \text{ReLU}!\left(\sum_k \alpha_k^c A^k\right)$$

• Uncertainty: if the top softmax probability is below a threshold $\tau$, the app returns "uncertain" instead of a label
• App: a Streamlit web interface so anyone can upload an image and get results instantly

Tech stack: Python, PyTorch, torchvision, Streamlit, NumPy, OpenCV, Matplotlib, GitHub

Results

Evaluated on a held-out test set, I focused on sensitivity (not missing sick patients) rather than accuracy alone:

$$\text{Sensitivity} = \frac{TP}{TP + FN}, \qquad \text{Specificity} = \frac{TN}{TN + FP}$$

• Sensitivity: [X%]
• Specificity: [X%]
• Accuracy: [X%]

Challenges I ran into

• Misleading accuracy: because the data is imbalanced, a model could look accurate while missing sick patients. I switched to sensitivity, specificity and a confusion matrix.
• Unreliable validation set: the provided validation split had only a handful of images, so I rebuilt the splits myself to get trustworthy evaluation.
• Overconfident predictions: neural networks can be confidently wrong, which is dangerous in healthcare. This motivated the uncertainty threshold.
• Checking the heatmaps: a good score isn't enough if the model is looking at the wrong thing (like text markers or borders). Grad-CAM let me catch this. [Describe anything you actually found.]
• [Add any real bug or setback you hit while building.]

What I learned

• Accuracy alone can be dangerous in medical AI. Choosing the right metric matters as much as choosing the right model.
• Explainability builds trust: seeing why a model decided helps spot errors and shortcuts.
• A responsible health tool should know its limits and say "I'm not sure."
• How to take a model from a notebook to a usable app that non-technical people can try.

Limitations and ethics

LungLens is a research prototype, not a medical device. It was trained on a single public dataset of pediatric X-rays and has not been clinically validated. It should never be used for real diagnosis without review by a qualified clinician.

What's next

• Test on more diverse datasets (different hospitals, ages, scanners)
• Add more conditions (e.g., tuberculosis, COVID-19)
• Calibrate confidence scores more rigorously
• Work with clinicians to evaluate real-world usefulness

Built With

• computer-vision
• convolutional-neural-networks
• deep-learning
• explainable-ai
• grad-cam
• machine-learning
• matplotlib
• medical-imaging
• numpy
• opencv
• pandas
• pillow
• python
• pytorch
• scikit-learn
• streamlit
• torchvision
• transfer-learning