Cardiomegaly Detection with EfficientNet & ViTs

Automating cardiomegaly detection can assist radiologists in early diagnosis, reducing workload and improving patient outcomes. AI-driven diagnostics can enhance accuracy, speed, and accessibility in medical imaging. 

For this project, I’m building a deep learning model to detect cardiomegaly (enlarged heart) from chest X-ray images, using the publicly available NIH ChestX-ray14 dataset. The dataset includes over 100,000 X-rays labeled for 14 different conditions, including cardiomegaly.

I'll be using transfer learning with state-of-the-art architectures such as EfficientNet, ResNet50, and Vision Transformers (ViT) to classify the images. These models will be fine-tuned on the dataset after applying preprocessing steps like resizing, normalization, and data augmentation.

To address class imbalance, techniques like weighted loss functions will be explored. Additionally, I’ll implement Grad-CAM to visualize which regions of the X-rays the model focuses on during prediction — an important step to ensure transparency and build trust in medical AI.

The goal is to create a high-performing and explainable model that could support radiologists in diagnosing cardiomegaly more accurately and efficiently, especially in high-volume or underserved clinical settings.