Journal Article

DKFZ-2024-02742

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Personalized deep learning auto-segmentation models for adaptive fractionated magnetic resonance-guided radiation therapy of the abdomen.

Kawula, M. ; Marschner, S. ; Wei, C. ; Ribeiro, M. F. ; Corradini, S. ; Belka, C.DKFZ* ; Landry, G. ; Kurz, C.

2025
AAPM College Park, Md.

Abstract: Manual contour corrections during fractionated magnetic resonance (MR)-guided radiotherapy (MRgRT) are time-consuming. Conventional population models for deep learning auto-segmentation might be suboptimal for MRgRT at MR-Linacs since they do not incorporate manual segmentation from treatment planning and previous fractions.In this work, we investigate patient-specific (PS) auto-segmentation methods leveraging expert-segmented planning and prior fraction MR images (MRIs) to improve auto-segmentation on consecutive treatment days.Data from 151 abdominal cancer patients treated at a 0.35 T MR-Linac (151 planning and 215 fraction MRIs) were included. Population baseline models (BMs) were trained on 107 planning MRIs for one-class segmentation of the aorta, bowel, duodenum, kidneys, liver, spinal canal, and stomach. PS models were obtained by fine-tuning the BMs using the planning MRI ( PS BM $\text{PS}_{\mathrm{BM}}$ ). Maximal improvement by continuously updating the PS models was investigated by adding the first four out of five fraction MRIs ( PS BM F4 $\text{PS}_{\mathrm{BM}}^{\operatorname{F4}}$ ). Similarly, PS models without BM were trained ( PS no BM $\text{PS}_{\mathrm{no BM}}$ and PS no BM F4 $\text{PS}_{\mathrm{no BM}}^{\operatorname{F4}}$ ). All hyperparameters were optimized using 23 patients, and the methods were tested on the remaining 21 patients. Evaluation involved Dice similarity coefficient (DSC), average ( HD avg $\text{HD}_{\rm avg}$ ) and the 95th percentile (HD95) Hausdorff distance. A qualitative contour assessment by a radiation oncologist was performed for BM, PS BM $\text{PS}_{\mathrm{BM}}$ , and PS no BM $\text{PS}_{\mathrm{no BM}}$ .PS BM F4 $\text{PS}_{\mathrm{BM}}^{\operatorname{F4}}$ and PS BM $\text{PS}_{\mathrm{BM}}$ networks had the best geometric performance. PS no BM $\text{PS}_{\mathrm{no BM}}$ and BMs showed similar DSC and HDs values, however PS no BM F4 $\text{PS}_{\mathrm{no BM}}^{\operatorname{F4}}$ models outperformed BMs. PS BM $\text{PS}_{\mathrm{BM}}$ predictions scored the best in the qualitative evaluation, followed by the BMs and PS no BM $\text{PS}_{\mathrm{no BM}}$ models.Personalized auto-segmentation models outperformed the population BMs. In most cases, PS BM $\text{PS}_{\mathrm{BM}}$ delineations were judged to be directly usable for treatment adaptation without further corrections, suggesting a potential time saving during fractionated treatment.

Keyword(s): MR‐Linac ; auto‐segmentation ; patient‐specific transfer learning

Note: 2025 Apr;52(4):2295-2304

---

Contributing Institute(s):

1. DKTK Koordinierungsstelle München (MU01)

Research Program(s):

1. 899 - ohne Topic (POF4-899) (POF4-899)

Appears in the scientific report 2024

---

Database coverage:
; ; Clarivate Analytics Master Journal List ; Current Contents - Clinical Medicine ; Current Contents - Life Sciences ; DEAL Wiley ; Ebsco Academic Search ; Essential Science Indicators ; IF < 5 ; JCR ; PubMed Central ; SCOPUS ; Science Citation Index Expanded ; Web of Science Core Collection

---