Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

One scale to rule them all: interpretable multi-scale Deep Learning for predicting cell survival after proton and carbon ion irradiation

Giulio Bordieri, Giorgio Cartechini, Anna Bianchi, Anna Selva, Valeria Conte, Marta Missiaggia, Francesco G. Cordoni

TL;DR

The paper tackles how energy deposition across nanometer, micrometer, and macroscopic scales governs cell survival and RBE in proton and carbon ion irradiation. It introduces an interpretable sequential-attention deep learning framework that fuses nanodosimetric, microdosimetric, and LET descriptors, trained on the PIDE dataset augmented with Monte Carlo simulations. The approach reveals scale-specific contributions—nanodosimetry often dominates at low doses while microdosimetry and LET contribute variably by endpoint and particle type—achieving competitive predictive accuracy (MAPE ~9.5% for $RBE_{10}$) and offering native interpretability via attention masks. These findings support a multiscale radiobiology perspective and point toward physics-informed, interpretable AI tools for particle therapy planning and optimization.

Abstract

The relationship between the physical characteristics of the radiation field and biological damage is central to both radiotherapy and radioprotection, yet the link between spatial scales of energy deposition and biological effects remains not entirely understood. To address this, we developed an interpretable deep learning model that predicts cell survival after proton and carbon ion irradiation, leveraging sequential attention to highlight relevant features and provide insight into the contribution of different energy deposition scales. Trained and tested on the PIDE dataset, our model incorporates, beside LET, nanodosimetric and microdosimetric quantities simulated with MC-Startrack and Open-TOPAS, enabling multi-scale characterization. While achieving high predictive accuracy, our approach also emphasizes transparency in decision-making. We demonstrate high accuracy in predicting RBE for in vitro experiments. Multiple scales are utilized concurrently, with no single spatial scale being predominant. Quantities defined at smaller spatial domains generally have a greater influence, whereas the LET plays a lesser role.

One scale to rule them all: interpretable multi-scale Deep Learning for predicting cell survival after proton and carbon ion irradiation

TL;DR

The paper tackles how energy deposition across nanometer, micrometer, and macroscopic scales governs cell survival and RBE in proton and carbon ion irradiation. It introduces an interpretable sequential-attention deep learning framework that fuses nanodosimetric, microdosimetric, and LET descriptors, trained on the PIDE dataset augmented with Monte Carlo simulations. The approach reveals scale-specific contributions—nanodosimetry often dominates at low doses while microdosimetry and LET contribute variably by endpoint and particle type—achieving competitive predictive accuracy (MAPE ~9.5% for ) and offering native interpretability via attention masks. These findings support a multiscale radiobiology perspective and point toward physics-informed, interpretable AI tools for particle therapy planning and optimization.

Abstract

The relationship between the physical characteristics of the radiation field and biological damage is central to both radiotherapy and radioprotection, yet the link between spatial scales of energy deposition and biological effects remains not entirely understood. To address this, we developed an interpretable deep learning model that predicts cell survival after proton and carbon ion irradiation, leveraging sequential attention to highlight relevant features and provide insight into the contribution of different energy deposition scales. Trained and tested on the PIDE dataset, our model incorporates, beside LET, nanodosimetric and microdosimetric quantities simulated with MC-Startrack and Open-TOPAS, enabling multi-scale characterization. While achieving high predictive accuracy, our approach also emphasizes transparency in decision-making. We demonstrate high accuracy in predicting RBE for in vitro experiments. Multiple scales are utilized concurrently, with no single spatial scale being predominant. Quantities defined at smaller spatial domains generally have a greater influence, whereas the LET plays a lesser role.
Paper Structure (14 sections, 16 equations, 8 figures)

This paper contains 14 sections, 16 equations, 8 figures.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. PIDE dataset
  4. Nanodosimetry
  5. Microdosimetry
  6. Deep Learning with attention mechanism
  7. Feature Transformer.
  8. Attentive Transformer and Feature Prior.
  9. Decision Steps and Aggregation.
  10. Sparse Regularization.
  11. Error analysis and biological endpoints considered
  12. Results
  13. Discussion
  14. Conclusions

Figures (8)

  • Figure 1: DL model workflow.
  • Figure 2: Model performance.
  • Figure 3: Global attention for the three endpoints considered.
  • Figure 4: Attention as a function of beam energy for protons and carbon ions for the three endpoints considered.
  • Figure 5: Spider plot with attention for protons and carbon ions for the three endpoints considered.
  • ...and 3 more figures