Augmentation Techniques to Improve Automatic Head and Neck Cancer Segmentation in Positron Emission Tomography Images
Kargar, Rahim (2025-11-18)
Augmentation Techniques to Improve Automatic Head and Neck Cancer Segmentation in Positron Emission Tomography Images
(18.11.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe20251124110674
https://urn.fi/URN:NBN:fi-fe20251124110674
Tiivistelmä
This M.Sc. thesis studies a deep learning-based approach for automatic segmentation of head and neck cancer (HNC) in positron emission tomography (PET) images using the U-Net architecture. The study includes a detailed analysis of PET imaging data from $89$ patients with confirmed positive HNC diagnoses, focusing on significant image slices identified through mask data and examining metrics such as the distribution of pixel intensities, the number of relevant slices processed per patient, and the performance of thresholding techniques for region identification. Segmentation accuracy is primarily evaluated using the Dice coefficient, to assess predictive performance and result reliability.
Recently, Liedes et al. \cite{Joonas} investigated the effectiveness of machine learning models in segmenting PET images. Specifically, when assessing true positive segmentations, these models achieved mean and median Dice scores of $0.79$ and $0.69$, respectively on test datasets of PET images, where the Image dimension is $128$, the threshold value is $0.25$, and the pixel limit is $5$. This thesis thoughtfully explores various cases concerning Image dimension, threshold values, and pixel limits, providing valuable insights for future research. Furthermore, we explore and refine augmentation techniques to enhance these results, including specific rotations, combinations of original and augmented training sets, and augmentations of training images and masks. Tests included applying random rotations between $-15$ and $15$, $90^{\circ}$ clockwise rotations, and techniques like horizontal flipping and Gaussian blurring.
For model validation, five cross-validation test sets were examined over five iterations each, producing a comprehensive evaluation with both mean and median Dice coefficients computed across all $25$ iterations. These final averaged metrics provide a summary of model performance across all the cases. This research highlights the promise of advanced U-Net-based models and tailored data augmentation strategies to improve PET image segmentation, supporting the broader integration of precise, automated cancer segmentation in clinical settings.
Recently, Liedes et al. \cite{Joonas} investigated the effectiveness of machine learning models in segmenting PET images. Specifically, when assessing true positive segmentations, these models achieved mean and median Dice scores of $0.79$ and $0.69$, respectively on test datasets of PET images, where the Image dimension is $128$, the threshold value is $0.25$, and the pixel limit is $5$. This thesis thoughtfully explores various cases concerning Image dimension, threshold values, and pixel limits, providing valuable insights for future research. Furthermore, we explore and refine augmentation techniques to enhance these results, including specific rotations, combinations of original and augmented training sets, and augmentations of training images and masks. Tests included applying random rotations between $-15$ and $15$, $90^{\circ}$ clockwise rotations, and techniques like horizontal flipping and Gaussian blurring.
For model validation, five cross-validation test sets were examined over five iterations each, producing a comprehensive evaluation with both mean and median Dice coefficients computed across all $25$ iterations. These final averaged metrics provide a summary of model performance across all the cases. This research highlights the promise of advanced U-Net-based models and tailored data augmentation strategies to improve PET image segmentation, supporting the broader integration of precise, automated cancer segmentation in clinical settings.