Quantitative Analysis of the Impact of Region of Interest Information on Deep Learning Algorithms for Thyroid Ultrasound Imaging

Goal: To quantitatively assess the impact of incorporating radiologist-defined Region of Interest (ROI) information in training deep learning models for thyroid ultrasound image classification and lesion localization. Methods: We compared a conventional convolutional neural network (CNN) trained without ROI information, interpreted through Grad-CAM for attention visualization, to Faster R-CNN and YOLOv2 models trained with radiologist-validated ROI masks. We also introduced an adapted mosaic-based composite input, derived from mosaic augmentation but implemented as fixed 1 × 2 and 2 × 2 layouts, to improve class balance and spatial diversity in training. Results: Models trained with ROI guidance achieved higher performance in both localization and classification compared to those trained without ROI. The average classification accuracy increased from about 80% in the baseline CNN to around 85% in ROI-guided models that shows an improvement of approximately 5 percentage points. The mean intersection over union between detected and radiologist-defined ROIs increased from approximately 33% to over 70%. The adapted mosaic input further stabilized performance across epochs and improved sensitivity while maintaining comparable specificity. Conclusions: Incorporating radiologist-defined ROI information and structured mosaic inputs significantly improves both diagnostic accuracy and localization precision. These results demonstrate that integrating ROI-guided learning with context-preserving composite inputs provides a reproducible framework for developing reliable AI systems in thyroid ultrasonography.