This work presents a detailed workflow for generating anonymized 3D head volumes from head and neck CT scans using the generation model architecture: Denoising Diffusion Probabilistic Model (DDPM). The primary objective is to develop a refacing technique that ensures patient anonymity, while maintaining essential anatomical details necessary for medical analysis. The workflow begins with the segmentation of the face region, isolating the area of interest within the 3D volume. The segmented face mask and the corresponding CT data are then pre-processed and used to train a 3D DDPM-based model, adapted for inpainting during the sampling. This model is designed to generate realistic volumes by selectively reconstructing regions defined by the face mask, ensuring seamless integration with the original data. The inpainting technique is crucial for achieving high-quality outputs that effectively anonymize the patient’s identity while retaining the structural integrity of the head and neck. Metrics to study the distribution of pixel intensities of the synthetic faces such as FID or MMD were computed depending on the depth of the diffuser. It allowed to sample the best realist synthetic faces. To evaluate the usability and effectiveness of the refaced data, a segmentation model was trained on the anonymized volumes. The results show similar performances as the model train on original data, demonstrating that the processed data remains suitable for further clinical analysis. Additionally, face recognition models were employed to assess the impact on anonymity.
Questo lavoro presenta un flusso di lavoro dettagliato per la generazione di volumi 3D anonimizzati della testa da scansioni TC della testa e del collo utilizzando un modello diffusionale probabilistico di denoising (DDPM). L'obiettivo primario è quello di sviluppare una tecnica di refacing che garantisca l'anonimato del paziente, pur mantenendo i dettagli anatomici essenziali necessari per l'analisi medica. Il flusso di lavoro inizia con la segmentazione della regione del viso, isolando l'area di interesse all'interno del volume 3D. La maschera facciale segmentata e i dati TC corrispondenti vengono quindi pre-elaborati e utilizzati per addestrare un modello 3D basato su DDPM, adattato per l'inpainting. Questo modello è progettato per generare volumi realistici ricostruendo selettivamente le regioni definite dalla maschera facciale, garantendo una perfetta integrazione con i dati originali. La tecnica di inpainting è fondamentale per ottenere risultati di alta qualità che anonimizzino efficacemente l'identità del paziente, mantenendo l'integrità strutturale della testa e del collo. Per valutare l'usabilità e l'efficacia dei dati ridipinti, è stato addestrato un modello di segmentazione sui volumi anonimizzati. I risultati mostrano prestazioni simili a quelle del modello addestrato sui dati originali, dimostrando che i dati elaborati rimangono adatti per ulteriori analisi cliniche. Inoltre, sono stati impiegati modelli di riconoscimento facciale per valutare l'impatto sull'anonimato.
A refacer for anonymization of CT scans using a denoised diffusion probabilistic model
Ducamp, Louis Marie Philippe
2023/2024
Abstract
This work presents a detailed workflow for generating anonymized 3D head volumes from head and neck CT scans using the generation model architecture: Denoising Diffusion Probabilistic Model (DDPM). The primary objective is to develop a refacing technique that ensures patient anonymity, while maintaining essential anatomical details necessary for medical analysis. The workflow begins with the segmentation of the face region, isolating the area of interest within the 3D volume. The segmented face mask and the corresponding CT data are then pre-processed and used to train a 3D DDPM-based model, adapted for inpainting during the sampling. This model is designed to generate realistic volumes by selectively reconstructing regions defined by the face mask, ensuring seamless integration with the original data. The inpainting technique is crucial for achieving high-quality outputs that effectively anonymize the patient’s identity while retaining the structural integrity of the head and neck. Metrics to study the distribution of pixel intensities of the synthetic faces such as FID or MMD were computed depending on the depth of the diffuser. It allowed to sample the best realist synthetic faces. To evaluate the usability and effectiveness of the refaced data, a segmentation model was trained on the anonymized volumes. The results show similar performances as the model train on original data, demonstrating that the processed data remains suitable for further clinical analysis. Additionally, face recognition models were employed to assess the impact on anonymity.File | Dimensione | Formato | |
---|---|---|---|
2024_10_Ducamp_Thesis_01.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: 2024_10_DUCAMP Master thesis : A refacer for anonymization of CT scans using a Denoised Diffusion Probabilistic Model
Dimensione
4.44 MB
Formato
Adobe PDF
|
4.44 MB | Adobe PDF | Visualizza/Apri |
2024_10_Ducamp_Executive Summary_02.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: 2024_10_DUCAMP Executive Summary : A refacer for anonymization of CT scans using a Denoised Diffusion Probabilistic Model
Dimensione
1.27 MB
Formato
Adobe PDF
|
1.27 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/227544