Automatic segmentation of brain tumors and stroke lesions from medical images using deep learning algorithms is crucial for prognosis, clinical assessment and treatment planning, and provides valuable clinical information. The analysis of the current state-of-the-art both for brain tumors and stroke lesions segmentations pointed out the most efficient techniques in these fields. It also underlined that most of them treat input MR images acquired with distinct modalities without caring of the divergence between the intensities with which the different cerebral and tumoral subregions are represented in these different modalities. Moreover, it was highlighted the almost complete absence of deep learning algorithms dealing with both brain tumors and stroke lesions segmentation. The main objective of this thesis was therefore to separate input images into the different available modalities, so that features extracted from images with divergent intensities may not be fused at earl levels, and to develop an Inter-pathology Learning technique between brain tumor and stroke lesion segmentation models, to transfer knowledge between those fields. The most promising and efficient model was identified in nnUNet, a self-adaptive framework that automatically adapts its network architecture to the specific task and dataset. Two different methods to separate input images of different MR modalities where thus developed: the first based on ensemble of models, while the second consisting in a multi-path network, modifying nnUNet by creating one different encoder for each input modality including dense connections, and resulting in the creation of Dense Multi-path nnUNet, which was the most promising one. The Dense Multi-path nnUNet models were then trained and evaluated using FeTS 2022 dataset for Brain Tumor Segmentation, while using ISLES 2022 dataset for Stroke Lesion Segmentation, being able to obtain Dice scores for the three tumoral regions (ED, NCR and ET) of 0.886, 0.823 and 0.903, with an average of 0.871; while the Dice score obtained for the segmentation of the stroke lesion was 0.660, overcoming the performances of nnUNet in both cases. An Inter-pathology Learning technique was also developed between the brain tumor segmentation model trained with FLAIR images, and the corresponding stroke lesion segmentation model trained with FLAIR images, showing superior performances of the basic model trained to segment stroke lesions.

Automatic segmentation of brain tumors and stroke lesions from medical images using deep learning algorithms is crucial for prognosis, clinical assessment and treatment planning, and provides valuable clinical information. The analysis of the current state-of-the-art both for brain tumors and stroke lesions segmentations pointed out the most efficient techniques in these fields. It also underlined that most of them treat input MR images acquired with distinct modalities without caring of the divergence between the intensities with which the different cerebral and tumoral subregions are represented in these different modalities. Moreover, it was highlighted the almost complete absence of deep learning algorithms dealing with both brain tumors and stroke lesions segmentation. The main objective of this thesis was therefore to separate input images into the different available modalities, so that features extracted from images with divergent intensities may not be fused at earl levels, and to develop an Inter-pathology Learning technique between brain tumor and stroke lesion segmentation models, to transfer knowledge between those fields. The most promising and efficient model was identified in nnUNet, a self-adaptive framework that automatically adapts its network architecture to the specific task and dataset. Two different methods to separate input images of different MR modalities where thus developed: the first based on ensemble of models, while the second consisting in a multi-path network, modifying nnUNet by creating one different encoder for each input modality including dense connections, and resulting in the creation of Dense Multi-path nnUNet, which was the most promising one. The Dense Multi-path nnUNet models were then trained and evaluated using FeTS 2022 dataset for Brain Tumor Segmentation, while using ISLES 2022 dataset for Stroke Lesion Segmentation, being able to obtain Dice scores for the three tumoral regions (ED, NCR and ET) of 0.886, 0.823 and 0.903, with an average of 0.871; while the Dice score obtained for the segmentation of the stroke lesion was 0.660, overcoming the performances of nnUNet in both cases. An Inter-pathology Learning technique was also developed between the brain tumor segmentation model trained with FLAIR images, and the corresponding stroke lesion segmentation model trained with FLAIR images, showing superior performances of the basic model trained to segment stroke lesions.

Analysis and adaptation of nnUNet (no-new-Unet) to process single modalities magnetic resonance images for brain tumors and ischemic stroke lesions segmentation

VIBERTI, ANDREA
2021/2022

Abstract

Automatic segmentation of brain tumors and stroke lesions from medical images using deep learning algorithms is crucial for prognosis, clinical assessment and treatment planning, and provides valuable clinical information. The analysis of the current state-of-the-art both for brain tumors and stroke lesions segmentations pointed out the most efficient techniques in these fields. It also underlined that most of them treat input MR images acquired with distinct modalities without caring of the divergence between the intensities with which the different cerebral and tumoral subregions are represented in these different modalities. Moreover, it was highlighted the almost complete absence of deep learning algorithms dealing with both brain tumors and stroke lesions segmentation. The main objective of this thesis was therefore to separate input images into the different available modalities, so that features extracted from images with divergent intensities may not be fused at earl levels, and to develop an Inter-pathology Learning technique between brain tumor and stroke lesion segmentation models, to transfer knowledge between those fields. The most promising and efficient model was identified in nnUNet, a self-adaptive framework that automatically adapts its network architecture to the specific task and dataset. Two different methods to separate input images of different MR modalities where thus developed: the first based on ensemble of models, while the second consisting in a multi-path network, modifying nnUNet by creating one different encoder for each input modality including dense connections, and resulting in the creation of Dense Multi-path nnUNet, which was the most promising one. The Dense Multi-path nnUNet models were then trained and evaluated using FeTS 2022 dataset for Brain Tumor Segmentation, while using ISLES 2022 dataset for Stroke Lesion Segmentation, being able to obtain Dice scores for the three tumoral regions (ED, NCR and ET) of 0.886, 0.823 and 0.903, with an average of 0.871; while the Dice score obtained for the segmentation of the stroke lesion was 0.660, overcoming the performances of nnUNet in both cases. An Inter-pathology Learning technique was also developed between the brain tumor segmentation model trained with FLAIR images, and the corresponding stroke lesion segmentation model trained with FLAIR images, showing superior performances of the basic model trained to segment stroke lesions.
2021
Analysis and adaptation of nnUNet (no-new-UNet) to process single modalities magnetic resonance images for brain tumors and ischemic stroke lesions segmentation
Automatic segmentation of brain tumors and stroke lesions from medical images using deep learning algorithms is crucial for prognosis, clinical assessment and treatment planning, and provides valuable clinical information. The analysis of the current state-of-the-art both for brain tumors and stroke lesions segmentations pointed out the most efficient techniques in these fields. It also underlined that most of them treat input MR images acquired with distinct modalities without caring of the divergence between the intensities with which the different cerebral and tumoral subregions are represented in these different modalities. Moreover, it was highlighted the almost complete absence of deep learning algorithms dealing with both brain tumors and stroke lesions segmentation. The main objective of this thesis was therefore to separate input images into the different available modalities, so that features extracted from images with divergent intensities may not be fused at earl levels, and to develop an Inter-pathology Learning technique between brain tumor and stroke lesion segmentation models, to transfer knowledge between those fields. The most promising and efficient model was identified in nnUNet, a self-adaptive framework that automatically adapts its network architecture to the specific task and dataset. Two different methods to separate input images of different MR modalities where thus developed: the first based on ensemble of models, while the second consisting in a multi-path network, modifying nnUNet by creating one different encoder for each input modality including dense connections, and resulting in the creation of Dense Multi-path nnUNet, which was the most promising one. The Dense Multi-path nnUNet models were then trained and evaluated using FeTS 2022 dataset for Brain Tumor Segmentation, while using ISLES 2022 dataset for Stroke Lesion Segmentation, being able to obtain Dice scores for the three tumoral regions (ED, NCR and ET) of 0.886, 0.823 and 0.903, with an average of 0.871; while the Dice score obtained for the segmentation of the stroke lesion was 0.660, overcoming the performances of nnUNet in both cases. An Inter-pathology Learning technique was also developed between the brain tumor segmentation model trained with FLAIR images, and the corresponding stroke lesion segmentation model trained with FLAIR images, showing superior performances of the basic model trained to segment stroke lesions.
Deep Learning
UNet
Segmentation
File in questo prodotto:
File Dimensione Formato  
Viberti_Andrea.pdf

Open Access dal 29/11/2023

Dimensione 7.7 MB
Formato Adobe PDF
7.7 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/39236