Work

Integration of clinical parameters and deep learning based ROI enhancement in Tc-99m PYP scintigraphy for predicting SPECT/CT-adjudicated Perugini categories in suspected transthyretin cardiac amyloidosis

Year: 2026

Type: article

Abstract: Technetium-99 m pyrophosphate (Tc-99 m PYP) planar scintigraphy is widely used for evaluating transthyretin cardiac amyloidosis (ATTR-CM). However, interpreting equivocal Perugini Grade 1 cardiac uptake remains a significant clinical challenge due to overlapping blood-pool activity and low signal-to-noise ratio. We aimed to develop and validate a multimodal deep learning framework to predict expert SPECT/CT adjudicated Perugini categories from planar imaging and pre-procedural clinical variables, with particular attention to Grade 1 examinations. In this multicenter study utilizing data from the Taiwan Multicenter TTCA-PYP Registry ( N = 438), we developed a multimodal neural network. The framework processes physician-verified cardiac planar ROIs (32 × 128 pixels) with a truncated Wide-ResNet50-2 encoder and 23 routine clinical variables with a frozen TabPFN embedding, aligned by a CLIP-style contrastive loss and fused by cross-attention, to predict SPECT/CT-adjudicated Perugini categories (Grade 0, Grade 1, Grades 2–3). Genetic testing status and study design were not inputs. Evaluation used 10 repetitions of a hospital- and grade-stratified assessment set with 10-fold cross-validation, decision curve analysis, and leave-one-center-out testing of the two centers with all grades (FEMH, NTUH). Overall accuracy was higher with the multimodal model than with imaging only (0.826 vs. 0.751; paired difference + 0.075, 95% CI 0.034 to 0.117), whereas macro-AUC was similar (0.925 vs. 0.917). In Grade 1, precision increased from 0.536 to 0.694 (difference + 0.159, 95% CI 0.056 to 0.261) and recall decreased from 0.704 to 0.683 (difference − 0.022, 95% CI − 0.120 to 0.077); F1-score and PR-AUC differences were less certain. Intervals used a Nadeau-Bengio correction for repeated resampling. With FEMH or NTUH held out, accuracy was 0.704 and 0.889, but Grade 1 recall was 0.312 in FEMH and Grade 0 recall 0.071 in NTUH. The multimodal framework may serve as an adjunct for standardizing expert imaging-based Perugini classification from planar Tc-99 m PYP examinations. It does not establish an etiologic diagnosis of ATTR-CM, and Grade 1 predictions remain indeterminate findings requiring monoclonal protein assessment and, where appropriate, tomographic or further clinical evaluation. Performance did not transfer uniformly to held-out centers, and independent validation is required.

Source: BMC Medical Imaging

DOI: 10.1186/s12880-026-03036-1