Preclinical deep learning framework shows improved dose prediction for unseen cancer sites

Photo by Navy Medicine / Unsplash

medRxiv Published April 24, 2026 Medically reviewed April 24, 2026 Study authors: Chang, H.-h.; Cardan, R.; Nedunoori, R.; Fiveash, J.; Popple, R.; Bodduluri, S.; Stanley, D. N.; Har… DOI ↗ By Dr. Lars van Dijk, PhD · Surgical, Procedural & Diagnostic

Key Takeaway

Consider this preclinical framework for dose prediction, but note its preliminary nature and lack of patient outcome data.

This is a preclinical development study of nnDoseNetv2, an auto-configured, end-to-end deep learning framework for dose prediction. The framework was trained on a single multi-site model integrating machine-specific beam geometry with 3D structural information, and its performance was compared to site-specific models.

The authors synthesized findings from 1,000 clinical plans. For sites seen during training (head and neck, prostate, breast, lung), performance was comparable to specialized site-specific models. For unseen sites, the framework outperformed site-specific models: for liver, the mean absolute error was 2.46% of prescription dose; for whole brain, the mean absolute error was 6.97% of prescription dose.

Key limitations noted by the authors include that this is not a clinical trial, performance is based on dose prediction accuracy rather than patient outcomes, generalizability to other cancer sites or treatment modalities was not evaluated, and details on model architecture and external validation are not provided in the abstract.

Practice relevance is restrained; the framework may enable scalable, high-fidelity radiotherapy dose prediction across diverse clinical scenarios, potentially reducing the need for site-specific model maintenance. However, certainty is low, as evidence is preliminary and based on a single dataset without external validation.

Study Details

EvidenceLevel 5

PublishedApr 2026

View Original Abstract ↓

Optimizing radiotherapy dose distributions remain a resource-intensive bottleneck. Existing AI-based dose prediction methods often have limited generalizability because they rely on small, heterogeneous datasets. We present nnDoseNetv2, an auto-configured, end-to-end framework for dose prediction across diverse disease sites (head and neck, prostate, breast, and lung), prescription levels (1.5-84 Gy), and treatment modalities (IMRT, VMAT, and 3D-CRT). By integrating machine-specific beam geometry with 3D structural information, the framework is designed to generalize across varied clinical scenarios. A single multi-site model was trained on 1,000 clinical plans. On sites seen during training, performance was comparable to specialized site-specific models. On unseen sites (liver and whole brain), the model outperformed site-specific models, with mean absolute errors of 2.46% and 6.97% of prescription, respectively. These results suggest that geometric awareness can bridge disparate anatomical domains while eliminating the need for site-specific model maintenance, providing a scalable and high-fidelity approach for personalized radiotherapy planning.