toraccic affections

This project explores the use of deep learning techniques for multilabel thoracic disease classification using chest X-ray images. The main objective was to build an initial medical imaging pipeline capable of detecting multiple pulmonary and thoracic conditions from radiographic studies. The project was developed using PyTorch and pretrained DenseNet architectures adapted for medical imaging tasks. Different experiments were conducted using grayscale chest X-rays resized to higher resolutions in order to preserve relevant anatomical information. The dataset includes multiple thoracic conditions such as Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Emphysema, Fibrosis, Hernia, Infiltration, Mass, Nodule, Pleural Thickening, Pneumonia, Pneumothorax, Pneumoperitoneum, Pneumomediastinum, Subcutaneous Emphysema, Tortuous Aorta and Calcification of the Aorta. The system was adapted from an initial binary classification approach into a multilabel classification pipeline using BCEWithLogitsLoss and sigmoid activation during inference. Several challenges related to medical datasets were addressed, including label alignment, class imbalance, tensor dimensionality issues and GPU optimization using CUDA in Google Colab. The project also incorporated evaluation methodologies more suitable for medical AI tasks, especially AUROC per pathology instead of relying exclusively on global accuracy metrics. This allowed a more realistic analysis of the model’s performance across different diseases. Additionally, exploratory work was performed to prepare the integration of explainability techniques such as Grad-CAM heatmaps to visualize the anatomical regions that contribute to the model predictions. This project represents a first practical exploration into AI-assisted radiology and medical computer vision, focusing on understanding the complete deep learning workflow for chest X-ray analysis rather than only optimizing benchmark performance.