Pneumonia Detection using Chest X-Rays

Pneumonia still kills hundreds of thousands of people a year, and the chest X-ray is the frontline diagnostic — yet reading one well takes a trained radiologist, and in many regions radiologists are scarce while imaging volumes keep rising. Computer-aided screening can help by triaging studies: flagging likely opacities for urgent review while clearly deferring the diagnosis to a human expert. This project builds that screening aid as a complete student system: a DenseNet-121 convolutional network fine-tuned on the public RSNA Pneumonia Detection Challenge dataset (26,684 frontal-view radiographs with radiologist-drawn bounding boxes around opacities), evaluated with per-class precision, recall and ROC-AUC on a held-out split, and wrapped in a study-viewer demo with Grad-CAM heatmaps, opacity boxes and a batch review queue. The report is explicit about what the model is — an educational pattern recognizer, not a medical device — and the demo carries that warning on every screen.

1. RSNA DICOM radiographs are converted to PNG, resized to 224×224 and normalized with ImageNet statistics.
2. The set is split into train/validation/test with stratification across the three classes; training images are augmented.
3. DenseNet-121 (ImageNet weights) has its classifier head replaced with a 3-way output and is fine-tuned with cross-entropy loss.
4. Every epoch logs training/validation loss, per-class precision/recall/F1 and ROC-AUC; the best checkpoint is kept.
5. At inference the viewer app preprocesses the study, runs the model, and renders class probabilities.
6. Grad-CAM backpropagates the predicted class score to the last convolutional layer to produce the heatmap overlay.

Tech stack:

• Python 3.10, PyTorch (DenseNet-121, transfer learning)
• Torchvision (transforms, augmentation, DataLoaders)
• pydicom + PIL (DICOM conversion and preprocessing)
• scikit-learn (per-class metrics, confusion matrix, ROC-AUC)
• Matplotlib, Seaborn (training curves, heatmaps, ROC plots)
• Flask viewer app (study viewer, batch queue, metrics dashboard)
• RSNA Pneumonia Detection Challenge dataset (Kaggle, 26,684 radiographs)

• [RSNA dataset pipeline] DICOM-to-PNG conversion, train/validation/test splits and augmentation (rotations, flips, contrast jitter) for the 26,684-image RSNA set, with the three-class taxonomy preserved.
• [DenseNet-121 classifier] ImageNet-pretrained DenseNet-121 fine-tuned for Normal / No Lung Opacity / Lung Opacity classification, with a training notebook logging loss, per-class metrics and ROC curves every epoch.
• [Grad-CAM explainability] Heatmap overlays generated from the final convolutional block show which lung regions drove each prediction, toggleable in the viewer.
• [Opacity bounding boxes] Predicted boxes with confidence scores localize suspected opacities, mirroring the radiologist annotations in the dataset.
• [Study viewer] Single-study analysis with overlay modes (original, heatmap, boxes, combined), prediction probabilities and DICOM metadata panel.
• [Batch review queue] A review grid across all three classes with filtering, so the demo walks through normal and abnormal studies.
• [Metrics dashboard] Confusion matrix, per-class precision/recall/F1 and one-vs-rest ROC curves computed by the notebook on your validation split.
• [Clinical-safety framing] Every screen carries the academic-prototype warning, and the report documents intended use, failure modes and why radiologist oversight is mandatory.

• Complete source code (dataset pipeline, training, evaluation, Grad-CAM, Flask viewer)
• Jupyter training notebook (buyer-run: preprocessing, fine-tuning, per-epoch metric logging)
• Project report PDF (background, dataset analysis, architecture, evaluation, error analysis, clinical-safety statement)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, Grad-CAM, ROC-AUC, medical-AI limits)
• Setup guide (environment, downloading the RSNA data from Kaggle, running training and the viewer)

• This is an academic prototype and is not intended to replace certified medical equipment, professional diagnosis or clinical decision-making — the report and every demo screen state this explicitly.
• The RSNA labels come from radiology reports with known label noise, and "No Lung Opacity" is a heterogeneous class — the model can inherit that ambiguity, which the error analysis discusses.
• Training covers adult frontal-view radiographs only; performance on pediatric, lateral or portable bedside films is unknown and not claimed.
• No measured accuracy is stated anywhere: the notebook computes per-class metrics on your own validation split during your build, and the report presents those as your build's results against the stated design target.
• The model sees pixels, not patients — it has no access to symptoms, history or lab values that a radiologist uses, so its output is at most a triage signal.