Medical Imaging · Deep Learning
Improved Pneumothorax Classification using DCGAN Augmentation
Explores DCGAN-generated chest X-rays for balancing SIIM training data. A fine-tuned ResNet-18 reached 98.6% training accuracy; Grad-CAM and t-SNE visualizations help inspect model attention and learned representations.
- 98.60% Train Accuracy
- 1.000 ROC AUC
Stack PyTorch / ResNet-18 / DCGAN / Grad-CAM
Problem
Pneumothorax, a collapsed lung, is diagnosable from chest radiographs but difficult to read even for experienced radiologists, and labeled positive cases are scarce relative to negatives. That class imbalance, combined with the barriers to sharing clinical data, makes it hard to train a classifier that generalizes rather than memorizing the majority class.
Approach
The project explores augmentation by synthesizing new positive examples. A deep convolutional GAN (DCGAN) is trained on the SIIM-ACR data to generate realistic synthetic pneumothorax X-rays, which are folded back into the training set to balance the positive and negative classes. A pretrained ResNet-18 classifier is then fine-tuned on the augmented data.
Grad-CAM visualizations help inspect which image regions influence predictions, while t-SNE visualizes class separation in the learned feature space.
Results
- 98.60% training accuracy
- ROC-AUC = 1.000
- Grad-CAM saliency maps and t-SNE embeddings for qualitative inspection of model attention and learned representations
These results describe the project experiments. Training accuracy and qualitative visualizations alone do not establish generalization or clinical usefulness; those claims require independent evaluation.