Optimizing thyroid nodule segmentation in thermal imaging with temporal sequences and advanced deep Learning backbones

Abstract

This study investigates methods to improve the segmentation of thyroid nodules in thermal imaging using deep learning models, with a focus on enhancing the Dice coefficient, a critical metric for model performance. We explore the integration of sequential image data through Long Short-Term Memory (LSTM) networks, hypothesizing that leveraging temporal features can significantly improve segmentation accuracy. Four novel deep learning models—U-Net with VGG16A, VGG16B, VGG19, and MobileNet backbones—were developed and evaluated. The research consisted of four studies: (1) analyzing the impact of sequence length (5, 10, and 20 images), which demonstrated a meaningful Dice improvement from 31.5 to 36.6 with longer sequences; (2) comparing feature-engineered versus raw data, revealing a tradeoff between sensitivity and precision; (3) assessing transfer learning approaches with VGG16 variants, where VGG16B achieved a 4 Dice improvement over VGG16A; (4) exploring alternative backbones (VGG19 and MobileNet) without substantial performance gains; and (5) conducting ablation experiments across all models and compared single-frame and multi-frame LSTM inputs. Sensitivity analysis highlighted the importance of reducing false negatives for better segmentation accuracy. Despite GPU and dataset limitations, our results indicate that LSTM-based sequential models significantly enhance segmentation performance, offering potential advancements in early thyroid nodule diagnosis and management. Future work will focus on multi-input designs and external validation to ensure generalizability and clinical applicability. © 2025 Elsevier Ltd