CheXpert Medical Imaging Competition

This project tackled automated anomaly detection in chest radiographs using over 100 thousand images from the CheXpert dataset,  a large-scale labeled dataset curated for real-world medical imaging research. Our goal was to predict the presence of 9 different conditions from chest X-rays (e.g., Cardiomegaly, Lung Opacity, Pleural Effusion) using supervised learning, submitting a CSV with probability predictions for each finding per image. Evaluation was based on mean squared error (MSE), scaled by class-wise variance.

​

Our team worked on the design, training, and evaluation of multiple deep learning models to identify optimal architecture-performance trade-offs:

• Vanilla CNN: Built and trained a baseline convolutional model to establish a performance benchmark.

• ResNet (Residual Networks): Implemented deeper architectures using skip connections (ResNet-18 and ResNet-50) to capture complex radiographic patterns.

• DenseNet: Leveraged densely connected CNNs to enhance feature propagation and minimize vanishing gradients in deeper layers.

​

To optimize model performance, we iteratively tuned:

• Learning rate schedules (step decay, cosine annealing)

• Optimizer selection (Adam vs. SGD with momentum)

• Batch size, weight decay, and dropout rates to reduce overfitting

• Loss functions (tested both BCEWithLogits and class-weighted variants due to label imbalance)

We also incorporated early stopping and model checkpointing to manage training stability and maximize generalization performance.

​

In terms of data processing and augmentation, we:

• Rescaled and normalized grayscale chest X-ray images to suit pretrained model input layers.

• Applied data augmentation techniques such as random horizontal flips, rotations, and brightness adjustments to improve robustness.

​

For the evaluation phase, we:

• Created scripts to compute scaled MSE across each of the 10 clinical labels.

• Visualized label-wise prediction distributions and investigated error trends across patient subgroups.

• Identified which conditions were most prone to false positives or false negatives and refined models accordingly.

​

This competition provided a hands-on opportunity to apply deep learning to a high-impact healthcare domain. We gained practical experience working with multi-label classification, medical image preprocessing, model performance debugging and optimization, and team-based ML pipeline development