Efficient Deep Learning Architecture for Detection and Recognition of Thyroid Nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Thyroid nodule diagnosis via ultrasonography presents challenges due to varied image features and blurred borders, complicating visual assessment.
• Deep learning (DL) shows promise in medical image analysis but faces hurdles in precision and efficiency for thyroid nodule recognition.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture, YOLOv3-DMRF, for enhanced detection and recognition of thyroid nodules in ultrasound images.
• To improve the accuracy and efficiency of differentiating between malignant and benign thyroid nodules.

Main Methods:

• Proposed a deep learning architecture, YOLOv3-DMRF, integrating a Dense Multireceptive Fields Convolutional Neural Network (DMRF-CNN) with multiscale detection layers.
• DMRF-CNN utilizes dilated convolutions with varying rates to preserve edge and texture features.
• Employed two datasets for training and validation, including augmented images and an open-access dataset of malignant and benign thyroid nodules.

Main Results:

• The YOLOv3-DMRF model demonstrated superior performance compared to state-of-the-art deep learning networks on both datasets.
• Achieved high mean average precision (mAP) values of 90.05% and 95.23% on the two test datasets.
• Exhibited efficient detection times of 3.7 and 2.2 seconds, respectively, indicating practical applicability.

Conclusions:

• The YOLOv3-DMRF deep learning architecture is highly efficient and accurate for the detection and recognition of thyroid nodules in ultrasound images.
• This approach offers a valuable tool to assist clinicians in the challenging diagnosis of thyroid nodules.
• The model's ability to handle diverse nodule appearances and sizes contributes to improved diagnostic precision.