The aim of this thesis is to develop an automatic segmentation system for the pancreas and pancreatic tumors based on 3D convolutional neural networks (nnU-Net). Specifically, two distinct models were implemented. The first model performs the automatic segmentation of the pancreas: trained on two datasets containing the pancreas ground truth, this model generates masks that accurately identify the organ in CT images. These masks are then used to extract a region of interest (ROI) corresponding to the pancreas. CT images are cropped based on this ROI, reducing the complexity of the problem and facilitating more accurate tumor segmentation. The second model, also based on nnU-Net, is trained exclusively on the cropped regions, focusing on the detection of pancreatic tumors. The two-stage approach enhances the accuracy and efficiency of the segmentation process, leveraging the strengths of 3D convolutional neural networks in handling large-volume CT scans. To highlight the computational impact of deep learning, comparative tests were conducted on a CPU (AMD EPYC 7B12, 4 cores) and a GPU (NVIDIA A100 40 GB). CPU times, extrapolated to the full training, are on the order of hundreds of days, whereas GPU training was completed in approximately 1 day and 13 hours for the TSM model and 2 days and 21 hours for the PSM model, with the CPU being over 300× slower than the GPU for this task.
L’obiettivo di questa tesi è lo sviluppo di un sistema di segmentazione automatica del pancreas e del tumore pancreatico basato su reti neurali convoluzionali 3D (nnU-Net). In particolare, sono stati realizzati due modelli distinti. Il primo modello provvede alla segmentazione automatica del pancreas: addestrato su un dataset che include la ground truth dell’organo, questo modello genera maschere che identificano con precisione il pancreas nelle immagini TC. Tali maschere vengono poi utilizzate per ricavare una regione d’interesse (ROI) corrispondente al pancreas, su cui vengono ritagliate le immagini TC per ridurre la complessità del problema e favorire una segmentazione più accurata della lesione. Il secondo modello, anch’esso basato su nnU-Net, è addestrato esclusivamente sulle regioni ritagliate, focalizzandosi sulla rilevazione del tumore pancreatico. L’approccio a due stadi consente di migliorare la precisione e l’efficienza del processo di segmentazione, sfruttando i vantaggi offerti dalle CNN 3D nella gestione di volumi TC di grandi dimensioni. Per mettere in evidenza l’impatto computazionale del deep learning, sono stati eseguiti test comparativi su CPU (AMD EPYC 7B12, 4 core) e GPU (NVIDIA A100 40 GB). I tempi misurati su CPU, proiettati sull’intero addestramento, risultano dell’ordine di centinaia di giorni, mentre su GPU l’addestramento effettivo è stato completato in circa 1 giorno e 13 ore per il modello TSM e 2 giorni e 21 ore per il modello PSM, con un rallentamento della CPU superiore a 300× rispetto alla GPU per questo task.
SEGMENTAZIONE AUTOMATICA DEL PANCREAS E DEL TUMORE DEL PANCREAS DA IMMAGINI DI TC
CRAPUZZI, GIOVANNI
2024/2025
Abstract
The aim of this thesis is to develop an automatic segmentation system for the pancreas and pancreatic tumors based on 3D convolutional neural networks (nnU-Net). Specifically, two distinct models were implemented. The first model performs the automatic segmentation of the pancreas: trained on two datasets containing the pancreas ground truth, this model generates masks that accurately identify the organ in CT images. These masks are then used to extract a region of interest (ROI) corresponding to the pancreas. CT images are cropped based on this ROI, reducing the complexity of the problem and facilitating more accurate tumor segmentation. The second model, also based on nnU-Net, is trained exclusively on the cropped regions, focusing on the detection of pancreatic tumors. The two-stage approach enhances the accuracy and efficiency of the segmentation process, leveraging the strengths of 3D convolutional neural networks in handling large-volume CT scans. To highlight the computational impact of deep learning, comparative tests were conducted on a CPU (AMD EPYC 7B12, 4 cores) and a GPU (NVIDIA A100 40 GB). CPU times, extrapolated to the full training, are on the order of hundreds of days, whereas GPU training was completed in approximately 1 day and 13 hours for the TSM model and 2 days and 21 hours for the PSM model, with the CPU being over 300× slower than the GPU for this task.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/70243