Cropped Tc-99m PYP images improved deep learning classification accuracy for ATTR-CM, with VGG11 achieving an overall AUC of 0.89 and Grade 1 AUC of 0.84.
Does a deep learning model using cropped Tc-99m PYP images improve the classification accuracy of ATTR-CM grades in patients with suspected ATTR-CM?
A deep learning model using cropped Tc-99m PYP images improves the diagnostic accuracy of distinguishing ATTR-CM grades, particularly the challenging Grade 1.
Absolute Event Rate: 0% vs 0%
Abstract Introduction Transthyretin amyloid cardiomyopathy (ATTR-CM) is an increasingly recognised cause of heart failure (HF) in older adults, with its prevalence rising due to advancements in diagnostics. While Tc-99m pyrophosphate (PYP) scintigraphy serves as a non-invasive alternative to biopsy, distinguishing between Grade 0 and Grade 1 remains challenging, often leading to misclassification and delayed treatment. This study proposes a deep learning approach leveraging cropped Tc-99m PYP images to enhance classification accuracy and facilitate early diagnosis. Purpose ATTR-CM is progressive and incurable condition. Although treatments such as tafamidis can slow disease progression, early and accurate diagnosis is crucial for timely intervention. However, differentiating between Grade 0 and Grade 1 Tc-99m PYP remains difficult, often resulting in diagnostic uncertainty. Standard 128×128 pixel Tc-99m PYP imaging has limitations in detecting Grade 1 ATTR-CM, highlighting the need for improved preprocessing and classification methods. This study investigates whether cropping Tc-99m PYP images to focus on the cardiac region can enhance deep learning-based classification, particularly in distinguishing Grade 1 from other grades. Methods This study analysed Tc-99m PYP scans from 383 patients in a multi-centre registry in Taiwan. The scans were classified into Grade 0 (n = 132), Grade 1 (n = 176), and Grade 2–3 (n = 75). The dataset was divided into an external test set (10%) and a cross-validation set (90%), with the latter further split into training (70%), validation (10%), and testing (10%) subsets. As part of the methodology, images were cropped based on radiologist-annotated heart centres, with a 16-pixel vertical extension to retain relevant cardiac structures while reducing background noise. Results This study compared multiple deep learning models and showed that cropped Tc-99m PYP images improve classification accuracy. Among the models evaluated, VGG11 achieved the highest accuracy, with an overall AUC of 0.89 and an AUC of 0.84 for Grade 1, outperforming existing methods. These findings highlight the effectiveness of ROI-based preprocessing in enhancing diagnostic accuracy and enabling earlier, more reliable detection of ATTR-CM. Conclusion Focusing on the cardiac region enhances model accuracy, particularly in distinguishing Grade 0 from Grade 1. Simpler CNN architectures outperformed complex models such as Vision Transformer and Swin Transformer, underscoring their suitability for small medical datasets. This study has several limitations, including dataset size, subjectivity in ROI definition, and real-world variability. Future research should refine cropping techniques, automate ROI detection, and incorporate domain-specific features to enhance clinical applicability.Fig 1.Schematic diagram of the method. Fig 2.Model performance comparison
Ou et al. (Sat,) reported a other. Cropped Tc-99m PYP images improved deep learning classification accuracy for ATTR-CM, with VGG11 achieving an overall AUC of 0.89 and Grade 1 AUC of 0.84.