Pathological diagnosis of thyroid nodules directly from ultrasonography by a weakly supervised deep learning framework

Category	Detail
Condition
Key Mechanisms	Deep learning framework for histological inference from ultrasound images, utilizing multi-instance learning and dual attention modules.
Target Population
Care Setting

ThyUS2Path framework demonstrated AUCs of 0.754 ± 0.035 and 0.735 ± 0.029 in five-fold cross validation, indicating moderate diagnostic performance.
Outperformed state-of-the-art MIL-based methods Maxpool and Meanpool, highlighting its effectiveness.
Achieved AUROCs of 0.70~0.80 and AUPRCs of 0.78~0.83 in external test sets, suggesting good generalizability.
Utilized 6014 images from 603 patients for training and 1978 images for validation, ensuring a robust dataset.
Addresses challenges in manual feature extraction and annotation in thyroid nodule assessment, improving diagnostic efficiency.

Regular follow-up imaging for nodules classified as TIRADS stage 3 or higher to assess changes over time.

Patients with thyroid nodules requiring histological diagnosis.

ThyUS2Path aims to enhance non-invasive diagnosis of thyroid cancer.

Utilize deep learning frameworks for improved consistency in ultrasound assessments, incorporating training for radiologists.
Implement multi-instance learning to manage varying nodule characteristics within patients, ensuring comprehensive training datasets.

Clinical Scorecard: Direct Histological Assessment of Thyroid Nodules via Ultrasonography Using a Weakly Supervised Deep Learning Approach