Modeling and Reversing Brain Lesions Using Diffusion Models
Omar Zamzam, Haleh Akrami, Anand Joshi, Richard Leahy · 2025
Read Modeling and Reversing Brain Lesions Using Diffusion Models by Omar Zamzam, Haleh Akrami, Anand Joshi, Richard Leahy (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05670.txt.
Introduction
Omar Zamzam, Haleh Akrami, Anand Joshi, and Richard Leahy present a model in two thousand twenty-five that reverses brain lesion growth in scans. Lesions mix truly damaged tissue with tissue that is merely pushed aside by swelling or growth. Standard segmentation labels both as one anomaly, losing that distinction.
Their diffusion model pipeline first segments abnormal regions in the MRI scan. It then estimates the deformation and restores displaced tissue to its place, isolating the core damage. Last, it fills in the core to estimate the healthy brain before the lesion. Since no real pre-lesion scans exist, they validated on synthetic lesioned brains from a forward growth model. Accuracy beat traditional methods.
Catalog record
| CREATOR | Omar Zamzam, Haleh Akrami, Anand Joshi, Richard Leahy |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-05670
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Modeling and Reversing Brain Lesions Using Diffusion Models Omar Zamzam, Haleh Akrami, Anand Joshi, Richard Leahy Abstract Brain lesions are abnormalities or injuries in brain tissue that are often detectable using magnetic resonance imaging (MRI), which reveals structural changes in the affected areas. This broad definition of brain lesions includes areas of the brain that are irreversibly damaged, as well as areas of brain tissue that are deformed as a result of lesion growth or swelling. Despite the importance of differentiating between damaged and deformed tissue, existing lesion segmentation methods overlook this distinction, labeling both of them as a single anomaly. In this work, we introduce a diffusion model-based framework for analyzing and reversing the brain lesion process. Our pipeline first segments abnormal regions in the brain, then estimates and reverses tissue deformations by restoring displaced tissue to its original position, isolating the core lesion area representing the initial damage. Finally, we inpaint the core lesion area to arrive at an estimation of the pre-lesion healthy brain. This proposed framework reverses a forward lesion growth process model that is well-established in biomechanical studies that model brain lesions. Our results demonstrate improved accuracy in lesion segmentation, characterization, and brain labeling compared to traditional methods, offering a robust tool for clinical and research applications in brain lesion analysis. Since pre-lesion healthy versions of abnormal brains are not available in any public dataset for validation of the reverse process, we simulate a forward model to synthesize multiple lesioned brain images. \midlauthor Ming Hsieh Department of Electrical and Computer Engineering University of Southern California Los Angeles, CA 90089, USA keywords Diffusion models, Registration, Lesion detection. 1 Introduction A brain lesion refers to an abnormality or injury within brain tissue, often detectable by imaging modalities such as magnetic resonance imaging (MRI), where it appears as alterations in the structure, function, or appearance of the brain. Lesions vary considerably in nature, and their definition encompasses any abnormalities that appear in brain tissue - this includes regions of irreversibly damaged tissue (such as necrotic zones) and areas of tissue deformation caused by mass effects from tumor growth, edema, or pressure from surrounding proliferating tissue. The extent of functional impairment correlates not only with the volume of the lesion but also with the degree of displacement or compression exerted on adjacent structures by these mass effects Sorribes et al. 2019 Greenspan 1972 . Edema, in particular, often manifests as swelling in the surrounding white matter, amplifying the overall volume expansion caused by the lesion, which may lead to further structural and functional damage. Necrotic tissue is indicative of irreversible damage, while adjacent deformed tissue can sometimes recover function, provided the compression or displacement is alleviated Wasserman et al. 1996 . Understanding the distinction between regions of permanently damaged tissue and those affected by potentially reversible deformation is paramount for accurate surgical planning. This differentiation enables the preservation of viable but deformed tissue while ensuring the removal of necrotic or irreversibly damaged regions. In addition, this understanding facilitates monitoring postoperative recovery and assessing the brain’s ability to heal. In this study, we propose a diffusion model-based framework for analyzing and reversing the brain lesion process. The proposed end-to-end framework is designed to first segment anomalous regions in the brain, then estimate tissue deformations within the anomalous regions. Estimating the tissue deformations allows for applying the inverse of these deformations to restore displaced tissue to its original position. Reversing deformed tissue back to the original position reduces the identified anomaly to only irreversible tissue damage. We inpaint the remaining damaged tissue to arrive at an estimation of the pre-lesion healthy brain shape. Our method offers several advantages: it segments brain lesions and distinguishes between irreversibly damaged regions and those deformed due to neighboring tissue growth or damage. It provides more robust and accurate labeling of lesioned brains compared to existing direct methods, aids in quantifying the impact of lesion growth on all brain regions of interest, and accurately estimates the brain’s pre-lesion structure. This proposed framework of reversing brain lesions assumes a forward process for lesion appearance and growth. This forward process is based on a model that is well-established in biomechanical studies that model brain lesions Sorribes et al. 2019 Greenspan 1972 Mohamed et al. 2006 Wasserman et al. 1996 where a brain region is damaged and then grows to deform and exert pressure on surrounding regions. To validate our proposed lesion reversal framework, we simulate the assumed forward model to allow for the synthesis of multiple lesioned brain images that serve as examples with known-groundtruth and that can be used to assess the accuracy of our proposed framework. 2 Related Work The general problems of lesion segmentation, intracranial pressure analysis in lesioned brains, and lesioned brain labeling have been extensively explored in the literature Kansal et al. 2000 Nalepa et al. 2019 Sorribes et al. 2019 Wyatt et al. 2022 . Early optimization-based methods provided holistic approaches to these problems Mohamed et al. 2006 , yet more recent learning-based methods have often examined these aspects separately. Given their ability to handle large datasets and provide fast processing, learning-based methods can offer complementary insights. By integrating biological models with statistical guidance from large datasets, these methods help address some of the limitations seen in purely model-based approaches, enabling more scalable solutions. Many developed methods tackle the separate problem of brain lesion segmentation. Those methods can be categorized into three categories: 1) Classical methods; where edge detection, thresholding, and template-based methods are used to segment lesions. 2) Machine Learning methods; which rely on feature extraction from images based on experts knowledge and training a machine learning model to detect the lesion features. 3) Deep Learning methods; where supervised learning is used to train deep neural networks to segment lesions using fully labeled datasets, or unsupervised learning can be used to denoise input images to get normal images at the output, and then thresholding the difference image comparing input and output. Jyothi and Singh 2023 provides a comprehensive review on state-of-the-art lesion segmentation methods. These methods mainly lack a clear way of understanding the lesions in the brain where restoration of functionality is unfeasible (damaged tissue) or feasible (deformed normal tissue) or understanding of the effects of the lesion on areas of the brain that surround the lesion. Lesioned brain labeling is a problem where the goal is to map information from a brain atlas to a lesioned brain image by learning a mapping between the atlas and the subject brain. This problem has been classically solved in the literature by directly registering lesioned brains to atlases (in the same way a healthy brain would be labeled) and mapping information from the atlas to the lesioned brain. These methods in deformable image registration often operate under the assumption of the existence of a diffeomorphism between the images being aligned Zou et al. 2022 ; Wu et al. 2022 ; Sotiras et al. 2013 ; Hill et al. 2001 , however, this assumption does not hold in medical scenarios such as pathological image to atlas registration or longitudinal studies comparing pre-operative and post-operative images Zacharaki et al. 2009 . The early work in Mohamed et al. 2006 provides an approach for the