-
Single Domain Generalization for Alzheimer's Detection from 3D MRIs with Pseudo-Morphological Augmentations and Contrastive Learning
Authors:
Zobia Batool,
Huseyin Ozkan,
Erchan Aptoula
Abstract:
Although Alzheimer's disease detection via MRIs has advanced significantly thanks to contemporary deep learning models, challenges such as class imbalance, protocol variations, and limited dataset diversity often hinder their generalization capacity. To address this issue, this article focuses on the single domain generalization setting, where given the data of one domain, a model is designed and…
▽ More
Although Alzheimer's disease detection via MRIs has advanced significantly thanks to contemporary deep learning models, challenges such as class imbalance, protocol variations, and limited dataset diversity often hinder their generalization capacity. To address this issue, this article focuses on the single domain generalization setting, where given the data of one domain, a model is designed and developed with maximal performance w.r.t. an unseen domain of distinct distribution. Since brain morphology is known to play a crucial role in Alzheimer's diagnosis, we propose the use of learnable pseudo-morphological modules aimed at producing shape-aware, anatomically meaningful class-specific augmentations in combination with a supervised contrastive learning module to extract robust class-specific representations. Experiments conducted across three datasets show improved performance and generalization capacity, especially under class imbalance and imaging protocol variations. The source code will be made available upon acceptance at https://github.com/zobia111/SDG-Alzheimer.
△ Less
Submitted 29 May, 2025; v1 submitted 28 May, 2025;
originally announced May 2025.
-
Distance Transform Guided Mixup for Alzheimer's Detection
Authors:
Zobia Batool,
Huseyin Ozkan,
Erchan Aptoula
Abstract:
Alzheimer's detection efforts aim to develop accurate models for early disease diagnosis. Significant advances have been achieved with convolutional neural networks and vision transformer based approaches. However, medical datasets suffer heavily from class imbalance, variations in imaging protocols, and limited dataset diversity, which hinder model generalization. To overcome these challenges, th…
▽ More
Alzheimer's detection efforts aim to develop accurate models for early disease diagnosis. Significant advances have been achieved with convolutional neural networks and vision transformer based approaches. However, medical datasets suffer heavily from class imbalance, variations in imaging protocols, and limited dataset diversity, which hinder model generalization. To overcome these challenges, this study focuses on single-domain generalization by extending the well-known mixup method. The key idea is to compute the distance transform of MRI scans, separate them spatially into multiple layers and then combine layers stemming from distinct samples to produce augmented images. The proposed approach generates diverse data while preserving the brain's structure. Experimental results show generalization performance improvement across both ADNI and AIBL datasets.
△ Less
Submitted 28 May, 2025;
originally announced May 2025.
-
Impact of alloy disorder on the band structure of compressively strained GaBiAs
Authors:
Muhammad Usman,
Christopher A. Broderick,
Zahida Batool,
Konstanze Hild,
Thomas J. C. Hosea,
Stephen J. Sweeney,
Eoin P. O'Reilly
Abstract:
The incorporation of bismuth (Bi) in GaAs results in a large reduction of the band gap energy (E$_g$) accompanied with a large increase in the spin-orbit splitting energy ($\bigtriangleup_{SO}$), leading to the condition that $\bigtriangleup_{SO} > E_g$ which is anticipated to reduce so-called CHSH Auger recombination losses whereby the energy and momentum of a recombining electron-hole pair is gi…
▽ More
The incorporation of bismuth (Bi) in GaAs results in a large reduction of the band gap energy (E$_g$) accompanied with a large increase in the spin-orbit splitting energy ($\bigtriangleup_{SO}$), leading to the condition that $\bigtriangleup_{SO} > E_g$ which is anticipated to reduce so-called CHSH Auger recombination losses whereby the energy and momentum of a recombining electron-hole pair is given to a second hole which is excited into the spin-orbit band. We theoretically investigate the electronic structure of experimentally grown GaBi$_x$As$_{1-x}$ samples on (100) GaAs substrates by directly comparing our data with room temperature photo-modulated reflectance (PR) measurements. Our atomistic theoretical calculations, in agreement with the PR measurements, confirm that E$_g$ is equal to $\bigtriangleup_{SO}$ for $\textit{x} \approx$ 9$%$. We then theoretically probe the inhomogeneous broadening of the interband transition energies as a function of the alloy disorder. The broadening associated with spin-split-off transitions arises from conventional alloy effects, while the behaviour of the heavy-hole transitions can be well described using a valence band-anticrossing model. We show that for the samples containing 8.5% and 10.4% Bi the difficulty in identifying a clear light-hole-related transition energy from the measured PR data is due to the significant broadening of the host matrix light-hole states as a result of the presence of a large number of Bi resonant states in the same energy range and disorder in the alloy. We further provide quantitative estimates of the impact of supercell size and the assumed random distribution of Bi atoms on the interband transition energies in GaBi$_{x}$As$_{1-x}$. Our calculations support a type-I band alignment at the GaBi$_x$As$_{1-x}$/GaAs interface, consistent with recent experimental findings.
△ Less
Submitted 5 March, 2013;
originally announced March 2013.
