Détection et classification de nodules thyroïdiens bénins et malins sur images échographiques par approches hybrides de Deep Learning

Abstract

The automatic detection and classification of thyroid nodules has seen significant development in recent years, thanks to rapid advances in medical imaging techniques and the emergence of deep learning methods, which have proven particularly effective in the field of computer vision. These methods are distinguished by their ability to automatically extract discriminating features from images, without requiring complex preprocessing steps, thus enabling the construction of rich hierarchical representations of visual information. This development has contributed to improving the efficiency of detection systems, making them better suited to the visual com- plexity that characterizes many medical pathologies, particularly in the field of ultrasound ima- ging diagnosis, such as the detection and classification of thyroid nodules. This is precisely what

we examined in our thesis.

We propose an automated system for detecting and classifying thyroid nodules into two cate- gories (benign/malignant) from ultrasound images. This system leverages deep learning archi- tectures combined using hybrid strategies to ensure high performance and satisfactory clinical robustness. We proposed five hybrid approaches based on the YOLOv8 and YOLOE models. These two models were initially trained individually. We also developed two cascade archi- tectures (YOLOv8→YOLOE and YOLOE→YOLOv8) and two prediction fusion approaches (NMS and WBF). Experiments were conducted on the ThyroidXL dataset, which comprises 11,545 ultrasound images from 4,093 patients. Individual evaluation shows that YOLOE achieves the best mAP@50 (90.1%) and YOLOv8 the best mAP@50-95 (61.1%), justifying their com- plementary nature. Among the hybrid approaches, the WBF+NMS fusion performs best overall, exhibiting the highest F1 score (92.9%) and the best recall for malignant targets (98.50%), re- flecting an optimal balance between reliability and comprehensiveness of detections. The WBF Fusion approach stands out with a very close malignant-class F1-score (89.52%) and a mAP@50 (97.32%), while the YOLOv8→YOLOE cascade achieves the best mAP@50 :95 (73.70%), in- dicating more precise spatial localization. Considering all the criteria, WBF+NMS is selected

as the optimal hybrid architecture.