ADVANCING AMYLOIDOSIS DIAGNOSIS: USING AI AND NON-SPECIFIC STAINING TECHNIQUES

Introduction: The diagnosis of amyloidosis traditionally requires Congo Red stained sections and light polarization, along with additional molecular characterization techniques. These steps complicate both the pre-analytic and analytic stages of diagnosis. Developing a system to streamline these processes could expedite diagnosis and reduce costs. AIM: To develop a computational model capable of classifying glomeruli affected by amyloidosis using routine stained sections without the need for light polarization, and to distinguish AL amyloidosis from AA amyloidosis without further staining. METHOD: A total of 4389 images of glomeruli from histological sections stained with H&E, PAS, or PAMS were utilized, comprising 374 glomeruli with amyloidosis confirmed by Congo Red staining, 1218 normal glomeruli, and 2797 with other lesions. To address dataset imbalance, data resampling, loss weighting, and voting techniques were employed. Various neural network architectures were tested, including VGG16, VGG19, Inception-ResNet, Inception, and Xception. RESULTS: The performance of the different network architectures was as follows, in terms of specificity and F1-score, respectively: VGG16 achieved 89.6% and 59.8%; VGG19 achieved 87.9% and 54.2%; Xception achieved 96.9% and 72.8%; Inception achieved 93.0% and 69.6%; Inception-ResNet achieved 96.4% and 78.0%. CONCLUSION: The model based on Inception-ResNet exhibited a specificity (96.4%) close to the best model (96.9%) but with a higher F1 score (78.0% versus 72.8%). This suggests potential for improved system performance. Future work should focus on training the system to classify cases of AA and AL amyloidosis more effectively.