Interpretable deep learning unlocks high-fidelity prediction for medical radioisotope production.

The efficient and pure production of medical radioisotopes, indispensable for diagnostic imaging and targeted radiotherapy, critically depends on accurate nuclear reaction cross section data. However, traditional physics-based models face limitations from inherent uncertainties and sparse experimental data, hindering optimal production strategies. Here, we introduce a comprehensive framework leveraging Bayesian-optimized deep neural networks to predict (p,2n) reaction cross sections for the production of clinically vital radioisotopes-⁴⁷Sc, ¹¹¹In, ¹²⁴I, and ¹⁶⁵Tm-with exceptional accuracy. Our models, trained and validated on evaluated data from the IAEA database, achieve a Pearson correlation coefficient R of 0.9997, significantly surpassing the performance of the TALYS-2.0 nuclear reaction code (R = 0.9783) using default parameters. Crucially, by integrating SHAP (SHapley Additive exPlanations) analysis, we provide unprecedented interpretability, elucidating the influence of projectile energy and nuclear structure features (notably neutron numbers of target and product nuclei) on cross-section predictions. This work demonstrates that data-driven, interpretable AI models can overcome long-standing challenges in nuclear data evaluation, offering a powerful tool to optimize cyclotron-based radioisotope production and advance nuclear medicine.