ETD

• autoencoder pretraining
• cam
• clinical data
• convolutional neural networks
• deep learning
• explainability
• gradcam
• heatmap
• hyperparameter optimization
• localization aware training
• medical image analysis
• optuna
• resnet50
• saliency may
• thyroid nodule classification
• ultrasound imaging

Accurate and interpretable classification of thyroid nodules from ultrasound imaging remains a critical challenge in computer-aided diagnosis, where both diagnostic performance and explainability are essential. This thesis proposes a two-stage deep learning framework that integrates autoencoder-based rep-
resentation learning, supervised classification with hyperparameter optimization, and a novel localization-aware training strategy driven by saliency supervision. Building on a ResNet-50 backbone, the system first leverages large-scale unsupervised pretraining to strengthen low-level feature extraction, fol-
lowed by a comprehensive Optuna search to identify an optimal classifier configuration. To enhance explainability and reduce spurious correlations, a saliency-aware loss function is introduced in the final training stage, encouraging Grad-CAM heatmaps to spatially align with radiologist-provided
bounding boxes of thyroid nodules.
Extensive experiments demonstrate that localization-aware training improves both predictive accuracy and interpretability. The final model achieves robust discrimination between benign and malignant nodules, while saliency maps become more anatomically focused and clinically meaningful. Quantitative explainability metrics (soft-IoU, pointing accuracy, energy-in-bbox) confirm that supervision of the saliency space consistently enhances alignment between model attention and ground-truth pathology. Qualitative results further show that the localization-aware model avoids the common pitfall of relying on artefacts, background texture, or acquisition-specific patterns.
Overall, this work demonstrates that explainable, localization-aware deep learning is a promising direction for thyroid nodule assessment, improving not only predictive performance but also transparency and clinical reliability. The proposed framework contributes a reproducible, modular pipeline that can serve as a foundation for future multimodal, clinically integrated AI systems in thyroid diagnostics.