Practical lessons from physician-led deep-learning prototyping in nuclear medicine imaging across increasing data...

Deep-learning applications in nuclear medicine are expanding, but many departments lack dedicated AI engineering support, creating a gap between interest and practical implementation. The aim of this study was to evaluate the feasibility of a physician-led, incremental workflow for developing reproducible deep-learning prototypes for nuclear medicine imaging across increasing data complexity (2D planar, sequential planar "3D-like", and true 3D DICOM) and to quantify prototype discrimination where reference standards were available.

Models were developed in Google Colab Pro using TensorFlow/Keras in a three-step framework. Step 1 used exported planar whole-body scintigraphy images (287 bone scintigraphy; 256 besilesomab scintigraphy) for radiopharmaceutical discrimination and an exploratory positive/negative subclassification without an external reference standard. Step 2 used ordered planar V/Q image stacks from 287 patients; reports provided reference-standard labels for acute pulmonary embolism (perfusion) and ventilation defects. Step 3 used true 3D PET DICOM volumes (46 [¹⁸F]FDG, 42 [¹⁸F]FDOPA, 42 [¹⁸F]FCH) with report-derived radiopharmaceutical labels. Datasets were split into 80% training and 20% validation subsets; performance was reported on the held-out validation subset unless otherwise stated. Primary outcome was feasibility (end-to-end pipeline completion and reproducible inference from an exported model); secondary outcomes included accuracy and AUC (plus sensitivity/specificity for Step 2).

Feasibility criteria were met across all three steps, with trained models exported and reloaded for inference in a separate application context. Step 1 radiopharmaceutical discrimination was successful in a small illustrative evaluation set (n = 20 images), while positive/negative subclassification demonstrated reduced reliability consistent with label ambiguity (exploratory). In Step 2, the perfusion model achieved an accuracy of 0.85 and an AUC of 0.96, with a sensitivity of 96.6% and a specificity of 86.6%; the ventilation model achieved an accuracy of 0.82 and an AUC of 0.93, with a sensitivity of 88.4% and a specificity of 76.9%. In Step 3, the 3D ResNet34 achieved a validation accuracy of 1.00 and an AUC of 0.95 for radiopharmaceutical discrimination.

This study demonstrates that deep-learning-based AI can be feasibly implemented even in small nuclear medicine departments. Although prototype in nature, it outlines a complete workflow from data extraction to model training and inference. Crucially, it highlights key limitations - such as reference-standard quality, task definition, preprocessing sensitivity, and the absence of an independent test set - that must be addressed to obtain a reliable and clinically meaningful in-house model.