AI-Augmented Thyroid Scintigraphy for Robust Classification
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912540323217408 |
|---|---|
| author | Sabouri, Maziar Hajianfar, Ghasem Sardouei, Alireza Rafiei Yazdani, Milad Asadzadeh, Azin Bagheri, Soroush Arabi, Mohsen Zakavi, Seyed Rasoul Askari, Emran Aghaee, Atena Wiseman, Sam Shahriari, Dena Zaidi, Habib Rahmim, Arman |
| author_facet | Sabouri, Maziar Hajianfar, Ghasem Sardouei, Alireza Rafiei Yazdani, Milad Asadzadeh, Azin Bagheri, Soroush Arabi, Mohsen Zakavi, Seyed Rasoul Askari, Emran Aghaee, Atena Wiseman, Sam Shahriari, Dena Zaidi, Habib Rahmim, Arman |
| contents | Purpose: Thyroid scintigraphy plays a vital role in diagnosing a range of thyroid disorders. While deep learning classification models hold significant promise in this domain, their effectiveness is frequently compromised by limited and imbalanced datasets. This study investigates the impact of three data augmentation strategies including Stable Diffusion (SD), Flow Matching (FM), and Conventional Augmentation (CA), on enhancing the performance of a ResNet18 classifier.
Methods: Anterior thyroid scintigraphy images from 2,954 patients across nine medical centers were classified into four categories: Diffuse Goiter (DG), Nodular Goiter (NG), Normal (NL), and Thyroiditis (TI). Data augmentation was performed using various SD and FM models, resulting in 18 distinct augmentation scenarios. Each augmented dataset was used to train a ResNet18 classifier. Model performance was assessed using class-wise and average precision, recall, F1-score, AUC, and image fidelity metrics (FID and KID).
Results: FM-based augmentation outperformed all other methods, achieving the highest classification accuracy and lowest FID/KID scores, indicating both improved model generalization and realistic image synthesis. SD1, combining image and prompt inputs in the inference process, was the most effective SD variant, suggesting that physician-generated prompts provide meaningful clinical context. O+FM+CA yielded the most balanced and robust performance across all classes.
Conclusion: Integrating FM and clinically-informed SD augmentation, especially when guided by expert prompts, substantially improves thyroid scintigraphy classification. These findings highlight the importance of leveraging both structured medical input and advanced generative models for more effective training on limited datasets. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_00366 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | AI-Augmented Thyroid Scintigraphy for Robust Classification Sabouri, Maziar Hajianfar, Ghasem Sardouei, Alireza Rafiei Yazdani, Milad Asadzadeh, Azin Bagheri, Soroush Arabi, Mohsen Zakavi, Seyed Rasoul Askari, Emran Aghaee, Atena Wiseman, Sam Shahriari, Dena Zaidi, Habib Rahmim, Arman Medical Physics Artificial Intelligence Computer Vision and Pattern Recognition Purpose: Thyroid scintigraphy plays a vital role in diagnosing a range of thyroid disorders. While deep learning classification models hold significant promise in this domain, their effectiveness is frequently compromised by limited and imbalanced datasets. This study investigates the impact of three data augmentation strategies including Stable Diffusion (SD), Flow Matching (FM), and Conventional Augmentation (CA), on enhancing the performance of a ResNet18 classifier. Methods: Anterior thyroid scintigraphy images from 2,954 patients across nine medical centers were classified into four categories: Diffuse Goiter (DG), Nodular Goiter (NG), Normal (NL), and Thyroiditis (TI). Data augmentation was performed using various SD and FM models, resulting in 18 distinct augmentation scenarios. Each augmented dataset was used to train a ResNet18 classifier. Model performance was assessed using class-wise and average precision, recall, F1-score, AUC, and image fidelity metrics (FID and KID). Results: FM-based augmentation outperformed all other methods, achieving the highest classification accuracy and lowest FID/KID scores, indicating both improved model generalization and realistic image synthesis. SD1, combining image and prompt inputs in the inference process, was the most effective SD variant, suggesting that physician-generated prompts provide meaningful clinical context. O+FM+CA yielded the most balanced and robust performance across all classes. Conclusion: Integrating FM and clinically-informed SD augmentation, especially when guided by expert prompts, substantially improves thyroid scintigraphy classification. These findings highlight the importance of leveraging both structured medical input and advanced generative models for more effective training on limited datasets. |
| title | AI-Augmented Thyroid Scintigraphy for Robust Classification |
| topic | Medical Physics Artificial Intelligence Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2503.00366 |