Automated Classification of Iron Accumulation in Patients with Alzheimer's Disease
Pennings, Hendrikus (2025-07-30)
Automated Classification of Iron Accumulation in Patients with Alzheimer's Disease
(30.07.2025)
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-fe2025080581142
https://urn.fi/URN:NBN:fi-fe2025080581142
Tiivistelmä
Alzheimer's disease (AD) is becoming one of the deadliest diseases for elderly people. Although considerable research is being conducted to gain a deeper understanding of AD, many questions about the pathology of the disease remain unanswered. One of these questions is about the role of iron in the disease. To evaluate the spreading pattern of iron in the brain, researchers histologically examine tissue sections from different brain regions of healthy and diagnosed brain donors post-mortem. Currently, they manually evaluate the extent of iron accumulation, but this process suffers from inter- and intra-variability among human experts and is time-consuming.
In this thesis, a method that includes a segmentation and classification model is proposed to automate the task of labelling histological images of brain regions. The goal of the segmentation model is to extract the grey matter from the image, as only the grey matter is relevant to this research. The result of the segmentation serves as the input for the classification model. The images are pre-processed beforehand, namely, they are cropped, downscaled, and the intensities are scaled. The proposed method uses the nnU-Net framework for the segmentation model. The produced segmentation masks are upscaled, and the grey matter is extracted from the original images. After that, the images are downsampled by a factor of 4, divided into patches, and a feature vector is created from each patch. The resulting feature vectors act as input for the classification model. The Clustering-constrained Attention Multiple Instance Learning (CLAM) framework is used for classification.
Our method yields satisfactory results for the segmentation model, outperforming human annotation, with an average Dice score of approximately 0.86 and an average NSD score of around 0.85. This means that the proposed method of using nnU-Net for the segmentation is usable as the input for the classification model. The classification model is unable to learn from the patches. It is inconclusive whether this is due to the experimental design or a software bug.
In this thesis, a method that includes a segmentation and classification model is proposed to automate the task of labelling histological images of brain regions. The goal of the segmentation model is to extract the grey matter from the image, as only the grey matter is relevant to this research. The result of the segmentation serves as the input for the classification model. The images are pre-processed beforehand, namely, they are cropped, downscaled, and the intensities are scaled. The proposed method uses the nnU-Net framework for the segmentation model. The produced segmentation masks are upscaled, and the grey matter is extracted from the original images. After that, the images are downsampled by a factor of 4, divided into patches, and a feature vector is created from each patch. The resulting feature vectors act as input for the classification model. The Clustering-constrained Attention Multiple Instance Learning (CLAM) framework is used for classification.
Our method yields satisfactory results for the segmentation model, outperforming human annotation, with an average Dice score of approximately 0.86 and an average NSD score of around 0.85. This means that the proposed method of using nnU-Net for the segmentation is usable as the input for the classification model. The classification model is unable to learn from the patches. It is inconclusive whether this is due to the experimental design or a software bug.