Poster Poster Program Therapy Physics

Explainable Deep Learning Framework Identifies Critical Low-Dose Sub-Regions for Symptomatic Pneumonia In NSCLC Radioimmunotherapy

Abstract
Purpose

Although consolidative immunotherapy following radiotherapy is standard for locally advanced non-small cell lung cancer (NSCLC), this combination increases the risk of ≥ Grade 2 symptomatic pneumonia (SP). This study utilized an explainable deep learning (DL) workflow to identify risk-associated dose sub-regions, thereby providing evidence to guide the future optimization of dose-volume constraints.

Methods

We retrospectively analyzed 315 NSCLC patients (training: 252; testing: 63) treated with combined radiotherapy and immunotherapy. The DL workflow integrated intrinsic and post-hoc explainability. For intrinsic explainability, CT images and dose distributions were segmented into six sub-regions based on dose intervals (5-10 Gy, and 10-60 Gy in 10 Gy increments). A dual-channel 3D ResNet was applied to extract features, a self-attention mechanism computed weights, and a multilayer perceptron was used to predict SP risk. For post-hoc explainability, Grad-CAM visualizations were generated and a novel activation cumulative histogram (AVH) was computed for sub-dose regions to analyze the intensity distribution of activation maps. Finally, restricted cubic spline (RCS) analysis (4 knots) was performed to determine dose-volume constraints.

Results

The model achieved AUCs of 0.726 (training cohort) and 0.724 (testing cohort). The 5-10 Gy sub-region received the highest attention weight (training: 0.447; validation: 0.446). Grad-CAM visualization revealed that the model primarily focused on low-dose areas to predict risk. The AVH indicated that, within the D5Gy-10Gy sub-region, an average of 14.63% of the area had attention values above 0.5. RCS analysis confirmed significant nonlinear associations between pneumonia risk and both V5Gy (P=0.018) and V10Gy (P=0.021), with risk plateauing at V5Gy ≥ 45% and V10Gy ≥ 38%.

Conclusion

This explainable DL workflow predicts SP risk, highlighting the critical importance of restricting low-dose lung volumes (V5Gy and V10Gy) to minimize toxicity in combined modalities.

People

Related

Similar sessions

Poster Poster Program
Jul 19 · 07:00
Python-Based Automation Framework for Annual Machine QA Data Archiving In Qatrack+

Annual water-tank measurements help ensure beam characteristics remain consistent with commissioning baselines. However, the lack of a standardized processing workflow and decentralized data storage makes it difficult to analyze...

Syed Bilal Ahmad, PhD
Therapy Physics 0 people interested
Poster Poster Program
Jul 19 · 07:00
User Expectations and Current Availability of HDR Brachytherapy Audits In Europe

The aim of this work was to evaluate the need to implement more dosimetric audits in high‐dose‐rate brachytherapy (HDR-BT) in Europe and to identify which characteristics such audits should meet according to users.

Javier Vijande, PhD Laura Oliver Cañamás
Therapy Physics 0 people interested