Multi-Class Classification of Thyroid Disorders Using Modified DenseNet-201 and Tc-99m Scintigraphy

Overview

This study presents a novel deep learning approach using a modified DenseNet-201 CNN model to classify seven thyroid disorders from Tc-99m scintigraphy images. Utilizing a real-world dataset of 1599 scans, the model addresses limitations of low-resolution imaging and manual interpretation, achieving improved diagnostic accuracy.

Background

The thyroid gland regulates metabolism through hormone production, and its dysfunction can lead to serious health risks. Common thyroid disorders include cold and hot nodules, goiters, thyroiditis, and toxic diffuse goiter, each varying in severity and treatment approach. Conventional diagnostic methods such as ultrasound, serum TSH measurement, and fine-needle aspiration biopsy have limitations in accurately distinguishing benign from malignant nodules. Tc-99m scintigraphy offers physiological assessment but suffers from low image resolution and subjective interpretation, motivating the use of AI-based techniques for improved diagnosis.

Data Highlights

Key Findings

• First detailed classification of seven thyroid disorders using scintigraphy images, expanding beyond prior binary or limited-class studies.
• Modified DenseNet-201 CNN model enhanced feature extraction from low-resolution Tc-99m scintigraphy images.
• Utilization of a large, real-world dataset from BINOR Cancer Hospital with moderate diversity and clinical relevance.
• Focus on nuclear medicine imaging aligns with real-world clinical workflows, avoiding reliance on high-resolution anatomical modalities.
• AI-based classification reduces inter- and intra-observer variability and mitigates diagnostic errors caused by manual image interpretation.
• Demonstrated potential to alleviate workload on nuclear medicine physicians and improve diagnostic consistency.

Clinical Implications

The integration of a modified DenseNet-201 CNN model for thyroid scintigraphy interpretation can enhance diagnostic accuracy and consistency in clinical practice. This approach supports early and precise identification of thyroid disorder types, enabling tailored treatment strategies and potentially reducing unnecessary interventions. Additionally, AI-assisted analysis may help address the shortage of nuclear medicine experts and reduce diagnostic fatigue.

Conclusion

This study demonstrates that deep learning applied to Tc-99m thyroid scintigraphy can effectively classify multiple thyroid disorders, overcoming limitations of conventional imaging and manual interpretation. The proposed model offers a promising tool to improve diagnostic workflows and patient outcomes in thyroid disease management.

References

1. Currie and Iqbal 2021 -- Comparison of scintigraphy and biochemical status for hyperthyroidism using ANN and CNN
2. Medhus et al. 2020 -- Exploration of CNNs for thyroid disorder detection
3. General references 1-29 -- Various studies on thyroid disorders, imaging, and AI applications