Metode hitre aproksimacije doznega polja visoke točnosti za fotonsko in protonsko radioterapijo : magistrsko delo

Škrijelj, Adis (avtor), Špiclin, Žiga (mentor)

Metode globokega učenja kot univerzalni aproksimatorji so lahko uporabni za pospešitev in izboljšanje točnosti izračuna doznega polja v radioterapiji. V tem delu smo razvili in vrednotili globoke modele, ki na podlagi hitro izračunanega doznega polja napovejo dozno polje z višjo točnostjo. Za hiter izračun doznega polja v primeru obsevanja s fotoni smo uporabili algoritem svinčnikastega snopa (ang. Pencil Beam - PB), medtem ko smo točnejšo dozo izračunali z algoritmom porušenega stožca (ang. Collapsed Cone Convolution - CCC). V primeru obsevanja z delci oziroma protoni smo za oba izračuna uporabili Monte Carlo (MC) simulacijo, kjer smo za hitri izračun doze uporabili 10000 delcev za simulacijo, medtem ko smo za izračun točnejše doze uporabili 1000000 delcev. Osnovna globoka arhitektura za preslikavo doznih polj je bila tipa U-Net za protone in kaskadni 3D U-Net (C3D) za fotone. Uporabili in vrednotili smo različne kriterijske funkcije, in sicer povprečno absolutno napako, napako s skrčenjem, zaznavno napako (ang. perceptual loss) in gradientno napako, s čimer smo želeli oceniti vpliv posamezne kriterijske funkcije na točnost izhodne napovedane doze. Poleg običajnih arhitektur smo za napovedovanje MC doznega polja protonov uporabili tudi U-Net z modulom pozornosti (ang. Attention U-Net – AU-Net) in generativno nasprotniško mrežo (ang. Generative Adversarial Networks - GAN), kjer smo za generator in diskriminator uporabili arhitekturo U-Net. Model GAN smo uporabili tudi za napovedovanje CCC doznega polja fotonov, z razliko, da smo za generator uporabili C3D arhitekturo. V splošnem smo za napovedovanje MC doznega polja protonov uporabili manj kompleksne modele z manjšim številom parametrov, saj smo kljub intenzivnejši regularizaciji in bogatenju podatkov med učenjem opazili, da je pogosto prihajalo do nasičenja učenja (ang. overfitting). Modele za napovedovanje CCC doznih polj fotonov smo učili in ovrednotili na dveh podatkovnih zbirkah, in sicer na interni zbirki primerov raka prostate ter odprtokodni zbirki OpenKBP za primere raka na območju glave in vratu. Modele za napovedovanje MC doznih polj protonov ovrednotili le na interni zbirki primerov raka prostate. Rezultati so pokazali, da je model C3D uspešno napovedal CCC dozna polja fotonov z visoko točnostjo. Na interni podatkovni zbirki je model dosegel povprečno absolutno napako 0,0753 Gy in 98,86 % povprečni delež ustreznosti gama indeksa pri kriteriju 1 % / 1 mm ter 99,95 % pri kriteriju 2 % / 2 mm. Na zbirki OpenKBP so bile vrednosti nekoliko nižje zaradi večje anatomske kompleksnosti primerov raka glave in vratu v primerjavi s prostato. Povprečna absolutna napaka je znašala 0,18 Gy, povprečni delež ustreznosti gama indeksa pa 90,5 % pri kriteriju 1 % / 1 mm ter 96,54 % pri kriteriju 2 % / 2 mm. To so bili najboljši rezultati, ki jih je dosegel C3D model na podlagi učenja z gradientno napako kot kriterijsko funkcijo. Model GAN je prav tako napovedal dozno polje razmeroma podobno referenčnemu CCC doznemu polju, vendar ne tako točno kakor prej omenjeni C3D model. Na OpenKBP podatkovni zbirki je povprečni delež ustreznosti gama indeksa pri kriteriju 1 % / 1 mm znašal 87,0 %, pri kriteriju 2 % / 2 mm pa 95,0 %. Pri napovedovanju MC doznega polja protonov je najbolj točne napovedi dosegal model U-Net s povprečno abolutno napako 0,17 Gy. Povprečni delež ustreznosti gama indeksa pri kriteriju 1 % / 1 mm je znašal 90,0 %, pri kriteriju 2 % / 2 mm pa 98,5 %. Prednosti predlaganih pristopov so visoka točnost, kratek čas izračuna in možnost uporabe v realnem času med postopkom načrtovanja radioterapije. Slabosti predstavljajo omejena velikost učne množice in potreba po nadaljnji validaciji na kliničnih podatkih različnih anatomskih regij. V prihodnje bi bilo smiselno vključiti večje in bolj heterogene zbirke, razširiti arhitekturo na difuzijske in/ali transformerske modele ter raziskati možnost sočasnega vključevanja kliničnih ciljev v proces učenja. Rezultati vrednotenja potrjujejo, da lahko modeli globokega učenja zanesljivo nadomestijo računsko zahtevne fizikalne algoritme ter omogočijo hiter izračun doznega polja ter tako prispevajo k avtomatizaciji in pospešitvi načrtovanja radioterapije.

• Jezik: Slovenski jezik
• Ključne besede: radioterapija, obsevanje, dozni izračun, globoko učenje, U-Net, magisteriji
• Vrsta gradiva: Magistrsko delo/naloga
• Tipologija: 2.09 - Magistrsko delo
• Organizacija: FE - Fakulteta za elektrotehniko
• Kraj izida: Ljubljana
• Založnik: A. Škrijelj
• Leto izida: 2026
• Št. strani: 1 spletni vir (1 datoteka PDF (XIV, 85 str.))
• PID: 20.500.12556/RUL-182460
• UDK: 615.849:004.8(043.3)
• COBISS.SI-ID: 281020163
• Datum objave v RUL: 12.05.2026

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Methods for fast high-accuracy dose field approximation for photon and proton radiotherapy : magistrski študijski program druge stopnje Elektrotehnika

Izvleček

Feasibility of applying deep learning methods as universal function approximators to accelerate and improve the accuracy and speed of dose field computation in radiotherapy was investigated. The principal objective was to develop and evaluate models capable of predicting a high-accuracy dose field from its low-accuracy. For photon irradiation, the Pencil Beam (PB) algorithm was employed for fast low-accuracy dose field calculation, while the Collapsed Cone Convolution (CCC) algorithm was used to obtain the reference high-accuracy dose field. For particle irradiation, both low- and high-accuracy dose field calculations were performed using Monte Carlo (MC) simulation, using 10000 and 1000000 particles for respective calculations of the input and the output (reference) dose fields. Standard model architectures were the U-Net and a Cascaded 3D U-Net (C3D) model for the MC based proton and CCC based photon dose field predictions, respectively. Several loss functions were evaluated; namely, mean absolute error, shrinkage loss, perceptual loss, and gradient loss in order to assess their individual impact on the accuracy of the predicted dose fields. In addition to the standard architectures, an Attention U-Net (AU-Net) and a Generative Adversarial Network (GAN), in which a U-Net served as both generator and discriminator were employed for the MC based proton dose field prediction. A GAN was also applied to CCC based photon dose field prediction, with the C3D model used as the generator. For MC based proton dose field prediction, the model architectures of lower complexity with fewer trainable parameters were adopted, as overfitting was observed despite extensive regularization and data augmentation. The CCC based photon dose field prediction models were trained and evaluated on two datasets. An internal prostate cancer dataset and the open source OpenKBP dataset comprising head and neck cancer cases. The MC based proton dose field prediction models were evaluated exclusively on the internal prostate cancer dataset. Obtained results demonstrated that the C3D model successfully predicted CCC based photon dose fields with high accuracy with respect to the reference values. On the internal prostate cancer dataset, the model achieved a mean absolute error of 0,0753 Gy, an average gamma pass rate of 98,86% under the 1 % / 1 mm criterion, and 99,95% under the 2 % / 2 mm criterion. Performance was somewhat lower on the OpenKBP dataset, attributable to greater anatomical complexity of the head and neck cancer cases as compared to the prostate cancer cases, with a mean absolute error of 0,18 Gy and average gamma index pass rates of 90,5% and 96,54% under the 1 % / 1 mm and 2 % / 2 mm criteria, respectively. These were the best results and were achieved using the C3D model trained with the gradient loss function. The GAN model also produced dose field distributions closely resembling the CCC based photon dose field reference, though with lower accuracy than the C3D model trained with the gradient loss. On the OpenKBP dataset of head and neck cancer cases, gamma pass rates were 87,0% and 95,0% under the 1 % / 1 mm and 2 % / 2 mm criteria, respectively. The U-Net model used for MC based proton dose field prediction achieved a mean absolute error of 0,17 Gy, with gamma index pass rates of 90,0 % and 98,5 % under the 1 % / 1 mm and 2 % / 2 mm criteria. The primary advantages of the proposed approaches include high predictive accuracy, short computation times, and potential applicability in real time treatment planning workflows. Limitations include the restricted size of the training datasets and the need for further validation on clinical data spanning diverse anatomical regions. Future work should consider incorporating larger and more heterogeneous datasets, extending the framework to diffusion and/or transformer based architectures, and exploring the integration of clinical objectives directly into the training process. Validation results confirm that deep learning methods can reliably replace computationally intensive physical dose calculation algorithms, enable fast dose field approximation, thereby contributing to the automation and acceleration of radiotherapy treatment planning.

• Ključne besede: radiotherapy, radiation, dose calculation, deep learning, U-Net