Morphology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using 3D MRIs
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Morphology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using 3D MRIs".
Title and authors: Tom: To recap, this paper is all about using mathematical morphology—the study of shapes—to create special ways to augment brain MRIs so a single AI model can detect Alzheimer's disease across different patient groups without needing separate training for each one.
Jane: That’s right. They propose this specific strategy where they combine these learnable shape manipulations with something called supervised contrastive learning to make sure the features the AI learns are really robust and not just noise from a specific scanner or protocol.
Tom: Exactly, and what's interesting is how they tailor these shape manipulations for each class. For Alzheimer's cases, they use erosion—which simulates losing brain tissue—and for the normal control group, they use dilation to simulate growth or healthy structure variations.
Jane: And for the milder cases, Mild Cognitive Impairment samples, they introduce some random shifts and scaling to create variation so the model learns to handle those transitional brain states better. It’s a very specific recipe depending on what kind of patient you are looking at.
Tom: The core mechanic is that this whole setup lets the AI see subtle structural differences that standard augmentation misses, which directly helps it generalize when it sees data from a completely new source.
Jane: That leads to the big question—when you look at the results on datasets like ADNI and AIBL, they show a clear improvement in performance compared to just using regular AI models without this shape-aware learning.
Tom: And the numbers backing that up are pretty solid too; they show drops in accuracy when you remove these morphological modules, meaning the structural information is actually important for getting those better scores.
Jane: It really changes how we think about training medical AI because instead of just feeding it pictures, we are actively teaching it to recognize specific anatomical patterns that define a disease.
Tom: This whole idea suggests that if we can successfully design these class-specific shape variations, the models will be way more reliable when deployed in real clinics where data quality isn't perfectly controlled.
Jane: It opens up a path for building AI tools that can adapt better to the messy, real-world data we get from different hospitals and scanners.
Tom: But it still has that caveat—the authors focused on training this on one source dataset and then testing it on others, so we have to be careful about how much those new target datasets differ from what they trained on <ref:2505.22465#pg4>.
Jane: That’s the balancing act there—getting high performance while acknowledging the need for more diverse training data down the road.
Tom: So basically, this is a sophisticated way to inject biological knowledge about brain structure directly into an AI's learning process to make it tougher against domain shift <ref:2505.22465#pg3>.
Jane: It’s a really smart direction for making medical AI more practical and reliable across different clinical settings.
The paper's summary: Tom: So, we’re looking at how they actually improved things in practice, moving beyond just the idea of using shape analysis to actually making those augmentations work better for the AI model.
Jane: Right, so they focused on a few key tweaks that make this whole system more effective than just throwing standard image transformations at the problem.
Tom: First up is these learnable pseudo-morphological modules themselves; instead of using fixed erosion or dilation, they make those operators adjustable based on what the model needs to see for each specific class.
Jane: That means the AI isn't just seeing a generic shape change, it’s seeing a shape change that is specifically relevant to whether the scan belongs to an Alzheimer's patient or someone with normal control features.
Tom: Then there’s the way they handle those MCI samples, where they use random affine transformations—those slight shifts and scaling—to create variation that bridges the gap between the healthy and diseased states.
Jane: And on top of all that, their supervised contrastive loss is working harder now because it’s being guided by these highly specific, learnable augmentations instead of generic ones.
Tom: The results show a big win there; they saw macro F1 scores jump significantly when they included this combination of the morphology and the contrastive learning together.
Jane: What that means for someone listening is that if you are building an AI for medical imaging, you can stop guessing what kind of image variations will help and start training the system with variations specifically engineered by shape analysis.
Tom: It shifts the focus from just getting a model to get better numbers on one dataset to designing a system that creates its own synthetic, class-specific challenges during training.
Jane: And this is really powerful because it directly tackles that problem of domain shift, which is usually where most AI tools fail when they move out of the lab and into a real hospital setting.
Tom: So the implication here is that future work needs to focus on making these morphological modules even more flexible so they can handle even more complex structural variations in brain MRIs.
Jane: Exactly, because while this works really well for ADNI and AIBL, testing it on something totally different requires those researchers to be very careful about how those shape rules translate across different brain types or scanner hardware.
The paper's improvements: Tom: So, to wrap up this discussion on "Morphology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using three dee MRIs," we’ve seen how they used learnable shape manipulations and contrastive learning to make the AI much tougher when it sees new types of brain scans.
Jane: It really shows that focusing on the actual physical structure of the brain, not just general image patterns, is a strong way to improve AI reliability for things like Alzheimer's detection.
Tom: Yeah, and the experimental results are pretty clear; they showed those performance boosts on ADNI and AIBL when they added these structural elements.
Jane: What this means for us is that we can build systems that are more robust when those scanners or protocols change because the AI has learned to focus on what actually matters structurally.
Tom: It’s a solid direction for making medical imaging AI more practical and reliable across different clinical settings, especially when dealing with the class imbalance issues they ran into.
Jane: I think the real value here is seeing how these mathematical operations can be integrated so seamlessly into a deep learning pipeline without losing that anatomical detail.
Tom: It’s definitely a step up from just using standard augmentations because those weren't targeting the disease-specific atrophy patterns as directly.
Jane: And even though they acknowledged it only trains on one source dataset, the way they used contrastive learning helps enforce class separation so that the features learned are more structured and less likely to get confused by noise.
Tom: The computational cost is manageable too; twenty-one point two million parameters isn't a massive overhead for the accuracy gains they achieved.
Jane: It’s a clever combination, Tom, using morphology to guide the data preparation and then contrastive learning to guide the representation learning itself.
Tom: Exactly. This paper gives us a concrete blueprint for how to inject domain knowledge about anatomy into deep learning models specifically for medical tasks like this one.
Jane: So, it sets a good benchmark for how we can use shape analysis to make AI more context-aware in high-stakes environments like neurology.
Conclusion: Tom: So we’ve been looking at "Morphology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using three dee MRIs." Basically, they took shape analysis and put it into an AI system to make it better at spotting Alzheimer's across different scans.
Jane: Right. They are using mathematical morphology to create specific ways of changing the MRI images—like erosion and dilation—so the AI learns features that are tied directly to brain structure, not just noise.
Tom: And they combine that with supervised contrastive learning, which is a technique used to make sure the AI’s internal features for different patients stay organized and separate.
Lu: What's interesting is how they tailor those shape modifications for each class; they use erosion for Alzheimer's and dilation for normal control samples because those operations visually intensify or alleviate the effects of brain atrophy.
Meng: I still have to wonder about the setup. If we have to design these specific shape modifications for every single class, isn't that going to become a massive manual headache in production?
Lalam: The paper says they combine these learnable modules with supervised contrastive learning to extract representations that are consistent across different environments, which helps them pull together the features of the same class while pushing apart those from different classes.
Jane: That makes sense because when you erode an image, it simulates losing brain tissue, which is exactly what we see in AD patients. But what about the Mild Cognitive Impairment group?
Tom: For the MCI group, they use random affine transformations—basically slight shifts and scaling—to introduce some variation so the model learns to handle that transitional state better.
Lu: The architecture itself is built on a pre-trained three dimensional U-Net encoder, but they drop the decoder part because they just want to use that encoder as a domain agnostic feature extractor, which then gets globally averaged into one compact feature embedding.
Meng: That sounds efficient for training on one dataset, but I have to ask about inference cost. How big is this model actually going to be when we deploy it in a clinical setting?
Lalam: The proposed method uses twenty-one point two million parameters and requires two thousand six hundred fifty point three gigapixels per forward pass for computation, which the authors say is a relatively modest overhead given the accuracy gains.
Tom: Okay, so they are using this U-Net backbone to get that feature embedding, and then they feed those features into a projection head that computes supervised contrastive loss against their learned augmentations.
Jane: That contrastive learning part is crucial because it’s not just about the augmentations; it’s about enforcing these representations to be very structured and discriminative, which should help separate the classes more cleanly than standard methods.
Wrap-up: Lu: The paper points out that existing domain generalization strategies, like those using distance transforms or auxiliary demographic tasks, often fail because they don't capture those specific disease-related anatomical changes in the brain that define Alzheimer's disease.
Tom: So their key improvement is explicitly modeling and learning from the morphology itself to create class-specific variations, rather than applying generic domain adaptation techniques.
Jane: That brings us to the overall conclusion they draw, which is that this combination of learnable morphological augmentations and supervised contrastive loss leads to superior generalization performance compared to baseline models on three different datasets: NACC, ADNI, and AIBL.
Meng: Superior performance is nice, but what about the limitations? The authors mentioned they are operating in a single domain setting initially, so I have to ask where this method might fall apart when we introduce truly novel imaging protocols that aren't covered by their training data.
Lalam: The paper does acknowledge that because it’s focused on single domain generalization, it assumes access to only one training dataset and no labels for the target domains, which is a limitation they have to manage.
Tom: So in summary, this paper introduces the Methodology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using three dee MRIs by combining learnable pseudo-morphological network modules with supervised contrastive learning to create class-specific augmentations that simulate brain morphology variations.
Jane: And the implication is that we can build models that are much more robust when we move from a lab setting where we train on one cohort, to a real clinic where the scanner or acquisition settings might be different.
Lu: If this works as well as they claim, it means future work could focus on refining those augmentation strategies further and testing this across even more diverse datasets beyond the three mentioned.
Meng: From a deployment standpoint, if we can get reliable performance like this without needing massive amounts of labeled target data, that significantly reduces the pipeline complexity for new deployments.
Lalam: The core message is that by being class-aware in our augmentations and using contrastive learning to enforce structure, we get representations that are much more stable when things change between training and testing domains.
Tom: So the final word is that this paper shows a way to use the shape analysis power of mathematical morphology to create robust, class-specific features for AD detection in three dee MRIs.
Jane: It’s a solid direction for anyone working on making medical imaging AI more practical and reliable across different clinical settings.
Tom: Next up, we’ll look at something completely different that deals with how large language models handle really long pieces of text and clinical questions.
Jane: We’ll be diving into "Investigating LLM Capabilities on Long Context Comprehension for Clinical Question Answering."
VPALab, Faculty of Engineering and Natural Sciences, Sabanci University
cs.CV
Submitted: 2025-05-28
Updated: 2026-11-17
Code: https://github.com/zobia111/SDG-Alzheimer
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Single Domain Generalization for Alzheimer’s Detection from 3D MRIs with Pseudo-Morphological Augmentations and Contrastive Learning is a new end-to-end single domain generalization (SDG) strategy…
Key concepts
- Single Domain Generalization (SDG)
- This is a technique where a model is trained on one set of data and then tested on completely new data from a different source. The goal is to make the model perform well even when the testing data has different characteristics, like variations in scanner hardware or acquisition protocols, without needing extra labeled data from those new sources.
- Pseudo-Morphological Augmentations
- These are synthetic image modifications based on mathematical morphology operators like erosion and dilation. Erosion simulates tissue loss (like atrophy in AD), while dilation simulates tissue gain, creating realistic shape variations that help the model learn robust features related to brain structure.
- Supervised Contrastive Learning
- This is a learning method used to refine the model's feature embeddings. It works by training the network to pull together features belonging to the same class (e.g., AD samples) while simultaneously pushing apart features from different classes, resulting in clearer and more distinct representations.
Terminology
Summary
Single Domain Generalization for Alzheimer’s Detection from 3D MRIs with Pseudo-Morphological Augmentations and Contrastive Learning is a new end-to-end single domain generalization (SDG) strategy specifically designed for Alzheimer’s disease detection that aims to produce class-specific augmentations that preserve a brain’s anatomical structure, by leveraging the inherent shape analysis capacity of mathematical morphology. p. 1 This strategy combines learnable pseudo-morphological network modules with supervised contrastive learning so as to simulate subtle brain morphology variations <ref:2505. p. 122465#pg4>.
Proposed SDG Approach
The proposed method combines a 3D U-Net encoder with class-specific 3D learnable morphological augmentations and supervised contrastive learning in order to improve AD detection under a realistic SDG setting <ref:2505.22465#pg11>. Access to a single training dataset is assumed, with no additional access to either target data or its labels <ref:2505. p. 222465#pg4>. During training, batches with only Mild Cognitive Impairment (MCI) samples undergo CutMix [34], a technique that swaps patches between images to increase intra-class diversity in addition to standard augmentations <ref:2505. p. 222465#pg11>. For the other two classes, AD and Normal Control (NC), 3D learnable morphological augmentations are computed via erosion and dilation respectively, motivated by the fact that these operators can visually intensify or alleviate the effect of brain atrophy <ref:2505. p. 222465#pg4>.
Model Architecture
The classification framework is built upon a pretrained 3D U-Net [35] <ref:2505.22465#pg12>. To adapt the architecture for the task under study, the decoder is dropped, and the encoder serves as a domain-agnostic feature extractor <ref:2505. p. 222465#pg12>. The encoder output is then globally average-pooled to produce a compact feature embedding <ref:2505. p. 222465#pg12>. Finally, a classifier head maps these features to class logits
Class-Specific Augmentations
To address distributional shifts between training and testing domains, class-specific learnable morphological 3D augmentation is incorporated <ref:2505.22465#pg12>. While standard augmentation techniques that apply identical transforms may improve robustness, they often miss disease-specific changes, potentially blurring class differences. p. 3 p. 3 The pseudo-Dilation Module (Dψ) generates synthetic NC samples by expanding regions of greater pixel intensity w. p. 3r.t. their surrounding via a learnable 3D dilation. p. 3 p. 3 The Pseudo-Erosion Module (Eψ) simulates brain tissue loss, similar to that observed in AD. p. 4 For MCI samples, random affine transformations are applied to introduce slight translation, scaling and contrast adjustments so as to produce an augmented sample. p. 4
Weighted Supervised Contrastive Learning
To encourage domain-invariant and class-discriminative representations, the features are provided as input to a projection head. The projected features are then used to compute supervised contrastive loss. p. 5 This loss pulls together representations of samples from the same class while pushing apart those from different classes, leading to more structured and discriminative feature embeddings. p. 6
Where the gains come from
The proposed method consistently achieved superior generalization performance compared to baseline approaches on both the ADNI and AIBL datasets. The improvements indicate enhanced domain-invariant feature learning and better separation between classes. p. 7 Ablation study results show that removing the morphological modules results in a drop in macro F1 score by 6. p. 85 percentage points on ADNI and 11. p. 88 percentage points on AIBL, showing the importance of structural information. p. 8 The loss helps the model separate the classes better, as removing it causes the F1 score to drop by 10. p. 82 percentage points on ADNI. p. 8
Computational Comparison
The proposed method uses 21.2 M parameters, requires 2065. p. 834 GFLOPs per forward pass, and consumes 494. p. 856 MB of memory. p. 7 This relatively modest overhead is justified by the significant accuracy and macro-F1 improvements shown in Tables 2 and 3. p. 9 The results demonstrate a favorable trade-off between computational cost and generalization performance. p. 9
Conclusion
This article introduced a new method to improve SDG performance for AD detection using 3D MRI data. By combining learnable 3D pseudo-morphological augmentations with a 3D U-Net encoder and supervised contrastive learning, our approach aimed to capture robust class-specific features. p. 9 Experimental results on the ADNI and AIBL datasets have shown superior performance over existing SDG techniques, especially when dealing with class imbalance and changes in imaging protocols. p. 9 Future work will aim to enhance computational efficiency, refine augmentation strategies, and evaluate the model across broader, more diverse datasets to improve generalization and clinical applicability. p. 9
REFERENCES
arXiv:2505.22465v1 [cs.CV] 30 May 2025
Single Domain Generalization for Alzheimer’s Detection from 3D MRIs with Pseudo-Morphological Augmentations and Contrastive Learning
Zobia Batool, Huseyin Ozkan, Erchan Aptoula
VPALab, Faculty of Engineering and Natural Sciences, Sabanci University, Istanbul, T¨urkiye
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 developed with maximal performance w. p. 0r.t. an unseen domain of distinct distribution. p. 0 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. p. 0 Experiments conducted across three datasets show improved performance and generalization capacity, especially under class imbalance and imaging protocol variations. p. 0
1 Introduction
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder affecting millions worldwide It is caused by a combination of factors and is characterized by gradual cognitive decline and structural brain changes [1]. Even though early and accurate diagnosis of AD is critical for timely intervention, it remains challenging due to the subtleness of the variations in brain morphology. p. 0 Magnetic resonance imaging (MRI) constitutes one of the key technologies used for the diagnosis of AD, yet detection models based on MRI datasets, often suffer from class imbalance issues and domain shift due to differences in scanner hardware and acquisition protocols, leading to poor generalization performance at deployment [2, 3, 4, 5]. p. 0
Among reported AD detection approaches, convnets have been investigated extensively through spatial and channel attention, frequency filtering, and optimization techniques [6, 7, 8]. p. 0 Whereas vision transformer based methods [9, 10, 11] have leveraged self-attention mechanisms to model long-range dependencies and integrate multimodal information. p. 0 Hybrid models combining convnet backbones with attention or gating mechanisms to balance interpretability and performance have also been studied [12, 13]. p. 0 However, even though these methods often report remarkable results on open-access datasets, they operate under the assumption of consistent training and testing distributions. p. 0 As such, their performance drops dramatically when confronted with real-world domain shifts, thus highlighting the need for generalization-focused approaches that capture disease-relevant patterns robustly across unseen domains. p. 0
Various strategies have been proposed to address domain shift in AD detection. p. 0 A recent study [14] introduced biologically informed priors using attention supervision, which helped improve interpretability. p. 0 Nevertheless, it depends on fixed saliency maps and does not explicitly model structural variations. p. 0 Moreover, approaches like Prototype-guided Multi-scale Domain Adaptation [15] and domain-knowledge-constrained models [16, 17], aim to address distribution shifts via multi-scale features or site-aware encoding. p. 0 But these methods mainly focus on global alignment rather than disease-specific structures and often overlook the progressive and localized brain atrophy patterns fundamental to AD diagnosis. p. 0
In this article, a new end-to-end single domain generalization (SDG) strategy is proposed, specifically designed for AD detection. p. 1 p. 1 It aims to produce class-specific augmentations that preserve a brain’s anatomical structure, by leveraging the inherent shape analysis capacity of mathematical morphology [18]. p. 1 p. 1 Specifically, the proposed approach combines learnable pseudo-morphological network modules with supervised contrastive learning so as to simulate subtle brain morphology variations. p. 1 p. 1 This strategy is designed to artificially increase intra-class variations through 3D learnable morphological augmentations, and thus enhance robustness to domain shift, allowing the model to generalize more effectively across datasets with different acquisition settings. p. 1 p. 1 It is built on a 3D U-Net backbone and is validated using three datasets (NACC, ADNI and AIBL), where it is shown to outperform baseline models. p. 1 p. 1
2 Related Work
2.1 AD Detection Approaches
Numerous studies have investigated AD detection using MRI data. A Multi-Attention-based Global 3D ResNet architecture is introduced in [6], that enhances feature representations using channel and spatial attention mechanisms. p. 1 It also contains a non-local block to capture long-range dependencies and outperforms other 3D convnets on the ADNI dataset. p. 1 Similarly, the 3D Global Fourier Network [7] uses global frequency filtering instead of spatial convolutions for long-range dependency modeling, achieving superior performance on both ADNI and AIBL datasets. p. 1 Building on attention mechanisms, the Dense Attention Network [12] combined convolutional layers with linear attention for strong classification and low parameter count. p. 1 Another study has introduced a multimodal surface-based transformer model [10] that integrated MRI and PET scans using a mid-fusion architecture. p. 1 It relies on self-attention and cross-attention blocks, outperforming volume-based baselines. p. 1 While these methods achieve impressive results, they often struggle to generalize across unseen domains, highlighting the need for DG. p. 1
2.2 Morphological Modules for AD Detection
Mathematical morphology has been extensively explored in the context of medical imaging [19, 20, 21] before the advent of deep learning due to its inherent capacity for shape analysis [22]. Consequently, multiple attempts have been made in order to implement its data driven counterpart, even though its non-linearity complicates differentiability and the back-propagation of errors via gradient descent. p. 1 For instance, one approach [23] introduced differentiable dilation and erosion using neural architecture search to integrate them into end-to-end models. p. 1 Another method [24] proposed multi-scale modules with learnable structuring elements. p. 1 A hybrid strategy [26] combined fixed morphological layers with shallow convnets, though it lacked flexibility and full end-to-end training. p. 1
More specifically in the context of morphology oriented AD detection, graph-theoretical metrics have been applied to morphological networks for early detection [27], while mesh-based graph convolutional networks have been used in [28] to model cortical structure. p. 1 Finally, a transformer-based model [29] incorporated morphologyaware augmentations but relied on external, non-learnable priors. p. 1 In contrast, our approach utilizes fully differentiable, class-aware pseudo-morphological modules for AD diagnosis. p. 1
In this article, a new end-to-end single domain generalization (SDG) strategy is proposed, specifically designed for AD detection. It aims to produce class-specific augmentations that preserve a brain’s anatomical structure, by leveraging the inherent shape analysis capacity of mathematical morphology [18]. Specifically, the proposed approach combines learnable pseudo-morphological network modules with supervised contrastive learning so as to simulate subtle brain morphology variations. This strategy is designed to artificially increase intra-class variations through 3D learnable morphological augmentations, and thus enhance robustness to domain shift, allowing the model to generalize more effectively across datasets with different acquisition settings. It is built on a 3D U-Net backbone and is validated using three datasets (NACC, ADNI and AIBL), where it is shown to outperform baseline models.
3 Proposed SDG Approach
The proposed method combines a 3D U-Net encoder with class-specific 3D learnable morphological augmentations and supervised contrastive learning in order to improve AD detection under a realistic SDG setting. Access to a single training dataset is assumed, with no additional access to either target data or its labels. p. 2 This combination of components encourages the model to generate more diverse, anatomically relevant variations while simultaneously enforcing representation consistency across domains. p. 2 During training, batches with only Mild Cognitive Impairment (MCI) samples undergo CutMix [34], a technique that swaps patches between images to increase intra-class diversity in addition to standard augmentations. p. 2 For the other two classes, AD and Normal Control (NC), 3D learnable morphological augmentations are computed via erosion and dilation respectively, motivated by the fact that these operators can visually intensify or alleviate the effect of brain atrophy. p. 2 Both original and augmented samples are processed through the encoder and a projection head, to compute contrastive embeddings. p. 2 The total loss combines weighted cross-entropy loss with inverse class frequency weights and supervised contrastive loss (Fig. p. 2 1).
3.1 Model Architecture
The classification framework is built upon a pretrained 3D U-Net [35]. To adapt the architecture for the task under study, the decoder is dropped, and the encoder serves as a domain-agnostic feature extractor. p. 2 The encoder output is then globally average-pooled to produce a compact feature embedding. p. 2 Finally, a classifier head maps these features to class logits. p. 2
3.2 Class-Specific Augmentations
To address distributional shifts between training and testing domains, class-specific learnable morphological 3D augmentation is incorporated. While standard augmentation techniques that apply identical transforms may improve robustness, they often miss disease-specific changes, potentially blurring class differences. To prevent this, the proposed method uses two learnable morphological modules [23] that approximate erosion and dilation with differentiable operations. p. 3 Erosion and dilation are the two fundamental operations of mathematical morphology, known for expanding image regions that are respectively darker and brighter w. p. 3r.t. their surroundings. p. 3 As such, when applied to a brain’s MRI sample, erosion’s effect becomes visually similar to intensifying brain atrophy (i. p. 3e. less brain matter). p. 3
3.2.1 Pseudo-Dilation Module (Dψ)
This module’s task is to generate synthetic NC samples, corresponding to healthy non-AD brain images. It achieves this by expanding regions of greater pixel intensity w. p. 3r.t. their surrounding via a learnable 3D dilation.
3.2.2 Pseudo-Erosion Module (Eψ)
This module is designed to simulate brain tissue loss, similar to that observed in AD. It works by expanding image regions darker than their surroundings, producing more severe-looking AD samples from existing ones. p. 4
3.2.3 Augmentations module for MCI (Tmci)
As a transitional state between NC and AD, MCI shows anatomical changes that may overlap with features of either class. Given an MCI sample, random affine transformations are applied in order to introduce slight translation, scaling and contrast adjustments so as to produce an augmented sample. p. 4
3.3 Weighted Supervised Contrastive Learning
To encourage domain-invariant and class-discriminative representati…
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Learnable pseudo-morphological modules that produce shape-aware, anatomically meaningful class-specific augmentations to simulate subtle brain morphology variations. This allows a model to artificially increase intra-class variations through 3D learnable morphological augmentations, enhancing robustness to domain shift across datasets with different acquisition settings.
-
Supervised contrastive learning module that extracts robust class-specific representations by pulling together representations of samples from the same class while pushing apart those from different classes. This results in more structured and discriminative feature embeddings, improving performance under class imbalance and protocol variations.
-
A hybrid training pipeline combining CutMix for Mild Cognitive Impairment (MCI) samples with 3D learnable morphological augmentations (erosion for AD, dilation for Normal Control) and supervised contrastive loss. This strategy encourages the model to generate more diverse, anatomically relevant variations while simultaneously enforcing representation consistency across domains.
-
A feature extractor built upon a pretrained 3D U-Net encoder that is designed to be domain-agnostic, producing a compact feature embedding via global average pooling before mapping to class logits. This forms the basis of the classification model and ensures that the core representation learning is independent of specific domain distributions during training on a single source dataset.
-
The system can detect Alzheimer’s disease from 3D MRI data with improved performance and generalization capacity, especially under class imbalance and imaging protocol variations, as demonstrated by superior macro F1 scores on the ADNI and AIBL datasets compared to baseline methods.
Sources
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs