DEEP LEARNING FOR COMPUTER VISION APPLICATIONS IN MEDICAL DIAGNOSTICS AND WILDLIFE MONITORING
Abstract
This dissertation explores innovative applications of deep learning and computer vision techniques across three distinct domains: medical imaging, dermatological diagnostics, and wildlife monitoring. The research addresses critical challenges in each field through the development and optimization of convolutional neural networks and other deep learning architectures.
The first study examines COVID-19 classification from X-ray images, comparing one-shot versus two-stage classification approaches using transfer learning with pre-trained models such as VGG16 and VGG19. The initial hypothesis was that breaking down the classification task into two optimized tasks would yield better results than one-shot classification. Results demonstrated that the single-stage approach achieved superior performance with 95% accuracy, contrary to initial hypotheses. The second study tackles the pervasive problem of class imbalance in dermoscopic image classification for skin cancer detection. Two approaches were systematically compared: traditional data augmentation techniques and Generative Adversarial Networks (GANs) for synthetic image generation. Fine-tuned deep learning models, including EfficientNet, ResNet50, Vision Transformers, and ConvNeXt, were evaluated on the augmented dataset and the synthetic dataset. Data augmentation proved more effective than GAN-generated synthetic images, with EfficientNet achieving 97.7% accuracy. The third study extends deep learning techniques to wildlife monitoring and conservation, developing a system for detecting and tracking beluga whales in aerial drone footage. The YOLOv7 object detection model achieved high precision and recall (92%–92% for adult belugas, 94%–89% for calves), while a novel post-processing algorithm improved multiple objects tracking accuracy from approximately 30% to 70%.
The collective findings advance the state of the art in applying deep learning to visual data across diverse domains, demonstrating effective solutions to key challenges in classification, class imbalance, and object tracking. This research contributes practical approaches that enhance diagnostic capabilities in healthcare, improve dermatological screenings, and support wildlife conservation efforts through more efficient monitoring techniques.