Pneumonia Detection using Transfer Learning

Pneumonia remains a leading infectious cause of death, and the chest X-ray is the most common study ordered for suspected cases. In high-burden settings the films pile up faster than radiologists can read them, and subtle opacities get missed when readers are fatigued. Automated assistance is a natural fit, but training a deep network from scratch on radiographs would need enormous labeled data. Transfer learning solves the data problem: a ResNet50 already trained on ImageNet has learned general visual structure, so fine-tuning its top layers on chest X-rays converges with tens of thousands of labeled films instead of millions. This project builds that pipeline end to end on the RSNA Pneumonia Detection Challenge set (26,684 frontal chest X-rays from the NIH Clinical Center, labeled Normal / No Lung Opacity-Not Normal / Lung Opacity), with Grad-CAM explanations so every prediction can be visually audited. The result is an assistive, explainable classifier whose limits are documented honestly, including the fact that it is an academic prototype and never a diagnostic device.

1. Chest X-ray images are loaded from DICOM, resized to 224x224 and contrast-enhanced with CLAHE to make lung detail consistent across scanners.
2. The dataset is split at the patient level so no patient's films appear in both training and validation.
3. A ResNet50 backbone initialized with ImageNet weights has its classifier head replaced with a 3-way output; the head trains first while the backbone stays frozen.
4. The top residual blocks are then unfrozen at a reduced learning rate for fine-tuning, with augmentation (rotation, flip, brightness) applied on the fly.
5. Each epoch logs training/validation loss and accuracy; the best checkpoint on validation is kept and exported.
6. At inference, the model returns class probabilities; Grad-CAM backpropagates the predicted class score to the last convolutional feature maps and overlays the resulting heatmap on the radiograph.
7. The demo app serves this pipeline behind a case gallery and explanation explorer for live demonstration.

Tech stack:

• Python 3.10, PyTorch with torchvision ResNet50
• OpenCV, pydicom (DICOM decoding), NumPy, pandas
• scikit-learn (patient-level splits, metrics)
• Matplotlib, Seaborn (training curves, confusion matrix, ROC)
• Flask demo app (case gallery, inference panel, Grad-CAM viewer)
• RSNA Pneumonia Detection Challenge dataset (26,684 frontal chest X-rays, NIH Clinical Center)

• [Transfer-learned classifier] ResNet50 pre-trained on ImageNet with its head replaced and top residual blocks fine-tuned on RSNA chest X-rays, in a two-phase schedule (head first, then deeper blocks at lower learning rate)
• [Three-class prediction] Outputs calibrated-style probabilities for Normal, No Lung Opacity / Not Normal, and Lung Opacity, with the winning class and confidence shown per radiograph
• [Grad-CAM explanations] Every prediction ships with an attention heatmap over the lung fields, generated from the final convolutional block, so the decision is visually auditable
• [Radiology-aware preprocessing] DICOM decode, resize to 224x224, CLAHE contrast enhancement and ImageNet normalization, with patient-level train/validation splits to prevent leakage
• [Case-gallery demo app] Web interface with sample cases, run-inference panel, Grad-CAM toggle and an explanation explorer (layer selector, opacity slider)
• [Training & evaluation dashboard] Epoch loss/accuracy curves, confusion matrix, ROC-AUC and per-class metrics, all logged by the notebook during the build
• [Augmentation pipeline] Rotation, horizontal flip and brightness jitter applied during training for robustness to acquisition variation
• [Viva-ready documentation] Report with data analysis, transfer-learning design choices, error analysis and the medical-prototype disclaimer stated explicitly

• Complete source code (data pipeline, training script, Grad-CAM module, demo app)
• Fine-tuned model weights exported from the included training run
• Jupyter training and evaluation notebook (buyer-run procedure: train, validate, evaluate)
• Project report PDF (background, dataset analysis, methodology, evaluation, error analysis, medical-prototype disclaimer)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, ResNet, Grad-CAM, metrics, limitations)
• Setup guide (environment, dataset download, training, running the demo)

• This is an academic research prototype, not a medical device: it must not be used for clinical diagnosis, and every prediction requires review by a qualified radiologist.
• No accuracy or sensitivity figure is promised: performance depends on the data split, preprocessing, hyperparameters and training conditions, and the report documents the measured results of the actual build.
• The model sees only frontal chest X-rays from the RSNA distribution; films from different scanners, pediatric cases or lateral views are outside its validated scope.
• Grad-CAM heatmaps highlight correlated regions, not causal pathology — they aid inspection, not diagnosis.
• The RSNA labels come from radiology reports, which carry their own noise; label quality bounds what any model can learn.