Transfer learning for medical image classification with PyTorch

The growing availability of pretrained deep learning models has significantly improved the accessibility and effectiveness of computer vision methods in specialized domains. Among these, medical image classification represents an important application area where transfer learning can provide strong performance even when data is limited. This project investigates the use of pretrained convolutional neural networks within a PyTorch pipeline for a three-class medical image classification task, with a particular focus on comparative evaluation, reproducibility, and practical experimentation.

The proposed workflow includes custom dataset construction, image preprocessing and augmentation, adaptation of pretrained models, GPU-aware training, and structured validation. The models considered in this study include ResNet50, VGG16, and GoogLeNet, all fine-tuned for the target classification problem. Their behavior is analyzed through multiple evaluation criteria, including accuracy, F1 score, recall, confusion matrices, and runtime. In addition, the project explores the influence of dataset size and validation strategy, incorporating techniques such as weighted sampling and Stratified K-Fold cross-validation to better address class imbalance and improve evaluation robustness.

Rather than focusing on the introduction of a novel architecture, this work aims to build a clear and reproducible experimental framework for comparing transferred CNN models on a domain-specific visual task. The repository therefore serves both as a deep learning project in medical imaging and as an example of practical machine learning workflow design implemented with PyTorch.