Deep learning based automatic segmentation of organs at risk in children with brain tumours
(2026) MSFT02 20262Medical Physics Programme
- Abstract
- The last decades have seen great improvement in the treatment of children with brain tumours,
5-year survival has increased from 20% in 1950 to 80% in 2020. Current treatment can consist
of a combination of surgery, chemotherapy and radiation therapy (RT). This improved chance
of survival has placed increased focus on the quality of life of the survivors post treatment.
RT increases the risk of neurocognitive complications, and the risk and complications depend
on the radiation dose to organs at risk (OAR) in the brain. Today, the OARs are manually
segmented for RT planning, which is time-consuming, since no commercially available solution
exists for automatic segmentation of this patient group. The aim of this project is to create... (More) - The last decades have seen great improvement in the treatment of children with brain tumours,
5-year survival has increased from 20% in 1950 to 80% in 2020. Current treatment can consist
of a combination of surgery, chemotherapy and radiation therapy (RT). This improved chance
of survival has placed increased focus on the quality of life of the survivors post treatment.
RT increases the risk of neurocognitive complications, and the risk and complications depend
on the radiation dose to organs at risk (OAR) in the brain. Today, the OARs are manually
segmented for RT planning, which is time-consuming, since no commercially available solution
exists for automatic segmentation of this patient group. The aim of this project is to create a
deep learning (DL) model that performs automatic segmentation of brain OARs in paediatric
patients.
A model was created based on a U-Net architecture using the framework no new U-Net (nnU-
net). The network takes a 3D image as an input and output segmentation masks of the OARs it
was trained on. nnU-Net was adapted to include residual connections on both the encoding and
decoding paths and CBAM attention modules on the encoding path. Tversky loss was added
to the compound loss function to improve performance on small OARs.
The training data consisted of 40 3D images, acquired using magnetic resonance imaging (MRI),
with manual segmentations, and 5 additional patients were placed in a test set to evaluate the
final model. The manual segmentations consisted of hypothalamus, thalamus, hippocampus,
brainstem, cerebellum, temporal- and frontal lobes. 5-fold cross-validation splits during training
were stratified based on age to reduce the anatomical effects of age. The final model was
evaluated with Dice’s similarity coefficient (DSC) and the 95th-percentile Hausdorff distance
(dH95) against the ground truth segmentations as well as a review by a physician to grade the
segmentations.
For cerebellum and brainstem, the model achieved median DSC and dH95 of 0.97 [0.84 - 0.97]
and 1.1 [1.0 - 5.0] mm, and 0.93 [0.79 - 0.94] and 2.0 [1.6 - 5.5] mm, respectively. The frontal
and temporal lobes showed median DSC and dH95 of 0.96 [0.88 - 0.96] and 2.0 [1.9 - 5.0] mm,
and 0.90 [0.71 - 0.92] and 4.4 [3.1 - 10.0] mm, respectively. For hippocampus, thalamus and
hypothalamus, the results were 0.86 [0.66 - 0.88] and 1.4 [1.1 - 5.0] mm, 0.86 [0.78 - 0.93] and
2.2 [1.6 - 4.0] mm, and 0.72 [0.50 - 0.74] and 2.2 [1.7 - 5.2] mm. The results from the physician
grading show almost identical overall scoring between the model and manual segmentation. The
model was graded worse on the frontal lobe as compared with the manual segmentation but
better on hippocampus.
Overall results show high DSC and low dH95 for all OARs, only hypothalamus having a no-
ticeably lower DSC and the temporal lobe having a higher dH95. nnU-Net provided a solid
foundation for the network and together with the modifications resulted in a strong architec-
ture. Evaluating the test patients, one outlier with considerably lower performance compared
to the other patients was observed. This patient had lower MRI image quality, indicating image
quality affect model performance. This is also reflected in the physician review as segmentations
of the lower quality image was graded lower with higher self-reported uncertainty. Setting the
caudal limit too high was the most stated flaw in the predicted segmentations, especially in the
frontal lobe. However, for the same patient both segmentations received almost equal grading.
The presented DL model shows great promise of reducing the workload for segmenting OAR
in paediatric patients with brain tumours, aiding in clinical treatment planning and improving
research possibilities. (Less) - Popular Abstract (Swedish)
- Under de senaste 50 åren har prognosen för barn med cancer i hjärnan förbättrats avsevärt.
År 1950 levde endast 20% av patienterna femår efter diagnostisering, idag är det över 80%. Dagens
behandling består av kirurgi följt av cellgifter och strålbehandling. Den förbättrade överlev-
naden har lagt ökat fokus på livskvaliteten hos barnen efter behandlingen. Strålbehandling
medför biverkningar och kan orsaka neurokoginitiva komplikationer, såsom intellektuell funk-
tionsnedsättning och försämrat minne senare i livet. Förståelse kring riskerna från strålbehandling
är viktigt för att minimera biverkningarna. Stråldos till olika delar av hjärnan medför olika
risker. En undersökning av riskerna kräver därför att hjärnstrukturerna... (More) - Under de senaste 50 åren har prognosen för barn med cancer i hjärnan förbättrats avsevärt.
År 1950 levde endast 20% av patienterna femår efter diagnostisering, idag är det över 80%. Dagens
behandling består av kirurgi följt av cellgifter och strålbehandling. Den förbättrade överlev-
naden har lagt ökat fokus på livskvaliteten hos barnen efter behandlingen. Strålbehandling
medför biverkningar och kan orsaka neurokoginitiva komplikationer, såsom intellektuell funk-
tionsnedsättning och försämrat minne senare i livet. Förståelse kring riskerna från strålbehandling
är viktigt för att minimera biverkningarna. Stråldos till olika delar av hjärnan medför olika
risker. En undersökning av riskerna kräver därför att hjärnstrukturerna identifieras och ritas in
i de bilder som används för att planera strålbehandlingen. Processen att rita in strukturerna
kallas segmentering och utförs av personal med särskild utbildning och är mycket tidskrävande,
vilket begränsar det tillgängliga underlaget för studier av biverkningar. Det finns i dagsläget
inget system som är optimerat för att segmentera bilder av hjärnan på barnpatienter.
Djupinlärning är en typ av AI som har visat sig vara mycket lämpat för inom många app-
likationer, bland annat för att automatisera segmenteringsprocessen. Djupinlärning använder
neurala nät som är inspirerat av hur den mänskliga hjärnan behandlar information i lager
av sammankopplade neuroner. För segmenteringsproblem används en djupinlärningsarkitektur
som kallas ett U-Net. Nätet tar in en bild som behandlas i lager där information om mönster
i bilderna extraheras. Modellen kan bara se en begränsad mängd pixlar i taget så efter varje
lager sänks upplösningen. Allt eftersom upplösningen minskar kan modellen se större mönster
i bilderna. När bilden är som mest komprimerad kan den generella positionen av en struktur
hittas. Bilden byggs sedan upp igen i lager, lika många som när den komprimerades, som nu
lägger till information i stället för att extrahera den. I slutet skapas en ny bild som tilldelar
varje pixel vilken aktuell struktur den tillhör eller om den utgör bakgrund.
En djupinlärningsmodell tränas på data som har ett facit, i detta fall bilder som redan har
manuellt inritade strukturer. Modellen lär sig då hur olika strukturer ska se ut baserat på
mönster i den inmatade bilden, likt hur en människa lär sig. Den data som användes i detta
projekt bestod av tredimensionella bilder tagna med magnetkamera för 45 barnpatienter som
behandlats med strålbehandling mot hjärntumör. Alla bilder har manuella segmenteringar
ritade av en neuropsykolog och godkända av läkare. Fem patienter åsidosattes för att kunna
testa den slutliga modellen på bilder modellen inte sett tidigare, vilket lämnar 40 patienter till
träning. Basen till modellen utgjordes av det populära ramverket no new U-Net som är designat
för att bygga segmenteringsmodeller. Vissa strukturer som modellen ska identifiera är väldigt
små, en ändring gjordes därför för att förbättra modellens förmåga att identifiera små mönster.
Modellen modifierades även för att bättre förstå vilka områden i bilden som den bör fokusera
på.
Modellen testades på de fem patienter som den inte sett under träningen. Resultaten visar på
att modellen ritar jämförbart med om olika läkare ritat samma struktur. Modellen presterar
likt det som väntas av kommersiella system som används på vuxna patienter, samt vad som
observerats i andra publicerade studier. Det finns därför goda förutsättningar för att den ska-
pade djupinlärningsmodellen kan användas för att automatisera inritning av hjärnstrukturer
för barnpatienter. Det möjliggör bättre studier på sambandet mellan stråldos till olika hjärn-
strukturer och de biverkningar barn upplever av strålbehandling, och i längden till förbättrad
strålbehandling med avseende på neurokognitiva biverkningar i framtiden. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9248511
- author
- Ewald, Erik
- supervisor
- organization
- course
- MSFT02 20262
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- language
- English
- id
- 9248511
- date added to LUP
- 2026-08-17 16:29:26
- date last changed
- 2026-08-18 11:14:58
@misc{9248511,
abstract = {{The last decades have seen great improvement in the treatment of children with brain tumours,
5-year survival has increased from 20% in 1950 to 80% in 2020. Current treatment can consist
of a combination of surgery, chemotherapy and radiation therapy (RT). This improved chance
of survival has placed increased focus on the quality of life of the survivors post treatment.
RT increases the risk of neurocognitive complications, and the risk and complications depend
on the radiation dose to organs at risk (OAR) in the brain. Today, the OARs are manually
segmented for RT planning, which is time-consuming, since no commercially available solution
exists for automatic segmentation of this patient group. The aim of this project is to create a
deep learning (DL) model that performs automatic segmentation of brain OARs in paediatric
patients.
A model was created based on a U-Net architecture using the framework no new U-Net (nnU-
net). The network takes a 3D image as an input and output segmentation masks of the OARs it
was trained on. nnU-Net was adapted to include residual connections on both the encoding and
decoding paths and CBAM attention modules on the encoding path. Tversky loss was added
to the compound loss function to improve performance on small OARs.
The training data consisted of 40 3D images, acquired using magnetic resonance imaging (MRI),
with manual segmentations, and 5 additional patients were placed in a test set to evaluate the
final model. The manual segmentations consisted of hypothalamus, thalamus, hippocampus,
brainstem, cerebellum, temporal- and frontal lobes. 5-fold cross-validation splits during training
were stratified based on age to reduce the anatomical effects of age. The final model was
evaluated with Dice’s similarity coefficient (DSC) and the 95th-percentile Hausdorff distance
(dH95) against the ground truth segmentations as well as a review by a physician to grade the
segmentations.
For cerebellum and brainstem, the model achieved median DSC and dH95 of 0.97 [0.84 - 0.97]
and 1.1 [1.0 - 5.0] mm, and 0.93 [0.79 - 0.94] and 2.0 [1.6 - 5.5] mm, respectively. The frontal
and temporal lobes showed median DSC and dH95 of 0.96 [0.88 - 0.96] and 2.0 [1.9 - 5.0] mm,
and 0.90 [0.71 - 0.92] and 4.4 [3.1 - 10.0] mm, respectively. For hippocampus, thalamus and
hypothalamus, the results were 0.86 [0.66 - 0.88] and 1.4 [1.1 - 5.0] mm, 0.86 [0.78 - 0.93] and
2.2 [1.6 - 4.0] mm, and 0.72 [0.50 - 0.74] and 2.2 [1.7 - 5.2] mm. The results from the physician
grading show almost identical overall scoring between the model and manual segmentation. The
model was graded worse on the frontal lobe as compared with the manual segmentation but
better on hippocampus.
Overall results show high DSC and low dH95 for all OARs, only hypothalamus having a no-
ticeably lower DSC and the temporal lobe having a higher dH95. nnU-Net provided a solid
foundation for the network and together with the modifications resulted in a strong architec-
ture. Evaluating the test patients, one outlier with considerably lower performance compared
to the other patients was observed. This patient had lower MRI image quality, indicating image
quality affect model performance. This is also reflected in the physician review as segmentations
of the lower quality image was graded lower with higher self-reported uncertainty. Setting the
caudal limit too high was the most stated flaw in the predicted segmentations, especially in the
frontal lobe. However, for the same patient both segmentations received almost equal grading.
The presented DL model shows great promise of reducing the workload for segmenting OAR
in paediatric patients with brain tumours, aiding in clinical treatment planning and improving
research possibilities.}},
author = {{Ewald, Erik}},
language = {{eng}},
note = {{Student Paper}},
title = {{Deep learning based automatic segmentation of organs at risk in children with brain tumours}},
year = {{2026}},
}