Imaging Tasks Research Articles

State-of-the-Art Deep Learning Methods for Microscopic Image Segmentation: Applications to Cells, Nuclei, and Tissues

Microscopic image segmentation (MIS) is a fundamental task in medical imaging and biological research, essential for precise analysis of cellular structures and tissues. Despite its importance, the segmentation process encounters significant challenges, including variability in imaging conditions, complex biological structures, and artefacts (e.g., noise), which can compromise the accuracy of traditional methods. The emergence of deep learning (DL) has catalyzed substantial advancements in addressing these issues. This systematic literature review (SLR) provides a comprehensive overview of state-of-the-art DL methods developed over the past six years for the segmentation of microscopic images. We critically analyze key contributions, emphasizing how these methods specifically tackle challenges in cell, nucleus, and tissue segmentation. Additionally, we evaluate the datasets and performance metrics employed in these studies. By synthesizing current advancements and identifying gaps in existing approaches, this review not only highlights the transformative potential of DL in enhancing diagnostic accuracy and research efficiency but also suggests directions for future research. The findings of this study have significant implications for improving methodologies in medical and biological applications, ultimately fostering better patient outcomes and advancing scientific understanding.

Spectral optimization using fast kV switching and filtration for photon counting CT with realistic detector responses: a simulation study.

Photon counting CT (PCCT) provides spectral measurements for material decomposition. However, the image noise (at a fixed dose) depends on the source spectrum. Our study investigates the potential benefits from spectral optimization using fast kV switching and filtration to reduce noise in material decomposition. The effect of the input spectra on noise performance in both two-basis material decomposition and three-basis material decomposition was compared using Cramer-Rao lower bound analysis in the projection domain and in a digital phantom study in the image domain. The fluences of different spectra were normalized using the CT dose index to maintain constant dose levels. Four detector response models based on Si or CdTe were included in the analysis. For single kV scans, kV selection can be optimized based on the imaging task and object size. Furthermore, our results suggest that noise in material decomposition can be substantially reduced with fast kV switching. For two-material decomposition, fast kV switching reduces the standard deviation (SD) by . For three-material decomposition, greater noise reduction in material images was found with fast kV switching (26.2% for calcium and 25.8% for iodine, in terms of SD), which suggests that challenging tasks benefit more from the richer spectral information provided by fast kV switching. The performance of PCCT in material decomposition can be improved by optimizing source spectrum settings. Task-specific tube voltages can be selected for single kV scans. Also, our results demonstrate that utilizing fast kV switching can substantially reduce the noise in material decomposition for both two- and three-material decompositions, and a fixed Gd filter can further enhance such improvements for two-material decomposition.

On estimating the effective sensitivity of a TOF-PET system for a particular imaging task

Comparing positron emission tomography (PET) systems which have different features is not straightforward. To address this, we propose to image the same object with all considered PET systems using a fixed scan time, and reconstruct from each scan an image at the same predefined spatial resolution. With such resolution matched reconstructions, the images should be identical except for their noise. Therefore, the PET system that produces the image with the lowest variance has the best performance. An analytical model is described to compute this variance, assuming that the PET system is (approximately) cylindrical, the imaged object is a uniform cylinder centered in the field of view, and the variance is only computed at the center of the reconstructed image. The model takes into account the solid angle covered by the detectors, the detector stopping power, the time-of-flight (TOF) resolution, the scatter fraction and the spatial resolution of the system, the attenuation and diameter of the cylinder and the desired spatial resolution of the reconstructed image. The inverse of this variance can be considered as the effective sensitivity of the system. This effective sensitivity can be calibrated based on the NEMA line source sensitivity. As a performance metric for scanning long objects, the minimum sensitivity achieved over the object is computed, with an optimal number of bed positions and optimal overlap between them. The influence of the spatial and TOF resolution on the effective sensitivity is verified with simulations.

Applications of cluster-based transfer learning in image and localization tasks

Transfer learning can address the issue of insufficient labels in machine learning. Using knowledge in a labeled domain (source domain) can assist in acquiring and learning knowledge in a domain (target domain) that lacks some or all labels. In this paper, we propose a new cluster-based semi-supervised transfer learning (CBSSTL) under a new assumption that samples in the target domain are unlabeled but contain cluster information. Furthermore, we propose a new transfer learning framework and a method for fine-tuning parameters. We tested and compared the proposed method with other unsupervised and semi-supervised transfer learning methods on well-known image datasets. The experimental results demonstrate the effectiveness of the proposed method. Additionally, we created a localization dataset for transfer learning. Finally, we tested and analyzed the proposed method on this dataset. Its particularly challenging nature makes it difficult for our method to work effectively.

SPW-TransUNet: three-dimensional computed tomography-cone beam computed tomography image registration with spatial perpendicular window Transformer.

Current medical image registration methods based on Transformer still encounter challenges, including significant local intensity differences and limited computational efficiency when dealing with three-dimensional (3D) computed tomography (CT) and cone beam CT (CBCT) images. These limitations hinder the precise alignment necessary for effective diagnosis and treatment planning. Therefore, the aim of this study is to develop a novel method that overcomes these challenges by enhancing feature interaction and computational efficiency in 3D medical image registration. This paper introduces a novel method that enhances feature interaction within Transformer by computing attention within resizable spatial perpendicular window (SPW). Additionally, it introduces a self-learning mapping control (SLMC) mechanism, which uses a mini convolutional neural network (CNN) to adaptively transform feature vectors into probability vectors. This approach is integrated into the UNet framework, resulting in the SPW-TransUNet. The effectiveness of the SPW-TransUNet is demonstrated through evaluations on two critical 3D medical imaging tasks: CT-CBCT registration and inter-CT registration. We utilized a range of evaluation metrics including Dice similarity coefficient (DICE), structural similarity index measure (SSIM), target registration error (TRE), and negative Jacobian percentage. The validation process involved comparative analysis against established baseline methods using statistical tests to ensure the robustness and reliability of our results. The proposed method demonstrated outstanding performance in the registration of 124 pairs of CT-CBCT lung images from 20 patients, achieving the lowest TRE of 2.16 mm and a minimal negative Jacobian of 0.126. It also recorded the highest SSIM and Dice coefficient of 86.87% and 88.28%, respectively. For the liver CT task involving 150 patients, the method achieved peak SSIM and DICE scores of 76.92% and 85.77%, respectively. Furthermore, ablation studies confirmed the effectiveness of the designed structural components. The SPW-TransUNet offers significant improvements in feature interaction and computational efficiency for medical image registration, providing an effective reference solution for patient and target localization in image-guided radiation therapy.

Real-time scan-free non-line-of-sight imaging

Real-time non-line-of-sight imaging is crucial for practical applications. Among existing methods, transient methods present the best visual reconstruction ability. However, most transient methods require a long acquisition time, thus failing to deal with real-time imaging tasks. Here, we provide a dual optical coupling model to describe the spatiotemporal propagation of photons in free space, then propose an efficient non-confocal transformation algorithm and establish a non-confocal time-to-space boundary migration model. Based on these, a scan-free boundary migration method is proposed. The data acquisition speed of the method can reach 151 fps, which is ∼7 times faster than the current fastest data acquisition method, while the overall imaging speed can also reach 19 fps. The background stability brought by fast scan-free acquisition makes the method suitable for dynamic scenes. In addition, the high robustness of the model to noise makes the method have the capability of non-line-of-sight imaging in outdoor environments during the daytime. To further enhance the practicality of this method in real-world scenarios, we exploit the statistical prior and propose a plug-in-and-play super-resolution method to extract higher spatial resolution signals, reducing the detector array requirement from 32 × 32 to 8 × 8 without compromising imaging quality, thus reducing the device expense of detectors.

State-of-the-Art Deep Learning CT Reconstruction Algorithms in Abdominal Imaging.

The implementation of deep neural networks has spurred the creation of deep learning reconstruction (DLR) CT algorithms. DLR CT techniques encompass a spectrum of deep learning-based methodologies that operate during the different steps of the image creation, prior to or after the traditional image formation process (eg, filtered backprojection [FBP] or iterative reconstruction [IR]), or alternatively by fully replacing FBP or IR techniques. DLR algorithms effectively facilitate the reduction of image noise associated with low photon counts from reduced radiation dose protocols. DLR methods have emerged as an effective solution to ameliorate limitations observed with prior CT image reconstruction algorithms, including FBP and IR algorithms, which are not able to preserve image texture and diagnostic performance at low radiation dose levels. An additional advantage of DLR algorithms is their high reconstruction speed, hence targeting the ideal triad of features for a CT image reconstruction (ie, the ability to consistently provide diagnostic-quality images and achieve radiation dose imaging levels as low as reasonably possible, with high reconstruction speed). An accumulated body of evidence supports the clinical use of DLR algorithms in abdominal imaging across multiple CT imaging tasks. The authors explore the technical aspects of DLR CT algorithms and examine various approaches to image synthesis in DLR creation. The clinical applications of DLR algorithms are highlighted across various abdominal CT imaging domains, with emphasis on the supporting evidence for diverse clinical tasks. An overview of the current limitations of and outlook for DLR algorithms for CT is provided. ©RSNA, 2024.

Second-order transformer network for video recognition

The video recognition community is undergoing a significant change in backbone shifting from CNNs to transformers. However, due to the temporal information existing in the video, vision transformers, which have been shown to be effective in image tasks, cannot model spatio-temporal structure in video recognition. In addition, current transformer-based video models exclusively utilize either class tokens or visual tokens for classification, failing to combine the merits of these two types of tokens. To solve the above problems, we propose a novel video second-order transformer network (ViSoT). As there are complex motion and appearance information in video, visual tokens of the last ViSoT layer are aggregated by cross-covariance pooling to model spatio-temporal information, which is combined with class tokens for classification. Meanwhile, a fast singular value power normalization is further introduced to achieve effective aggregation of visual tokens. For temporal modeling, before ViSoT block temporal convolutions are performed continuously in the early stage after the convolution stem via a short-cut connection. In ViSoT, token shift module and space–time attention are proposed for modeling temporal relations and spatio-temporal interaction across adjacent frames respectively. Validating ViSoT on four video benchmarks demonstrates its effectiveness and efficiency, comparable to or better than state-of-the-art methods with fewer parameters and GFLOPs.

Multi-consistency for semi-supervised medical image segmentation via diffusion models

Medical image segmentation presents a formidable challenge, compounded by the scarcity of annotated data in numerous datasets. Semi-supervised methods offer viable solutions to mitigate that, while the image generation capability of diffusion models has shown potential in capturing semantically meaningful information. This paper introduces Multi-Consistency for semi-supervised medical image Segmentation via Diffusion Models (MCSD). We propose a Diffusion-based Feature-guided Module (DFM) that extracts features from pre-trained diffusion models and uses multi-scale features to guide multi-consistency segmentation networks. Additionally, we introduce the Dual-branch Image Consistency (DIC) strategy, which performs multi-consistency learning by generating two independent strong augmented images and optimizing the network by encouraging consistency between strong and weak inputs at multiple levels of features and images. Our method outperforms current approaches, demonstrating its effectiveness in semi-supervised medical image segmentation through experimental results for various labeled data ratios. Furthermore, this work points out the potentiality of diffusion models in semi-supervised medical image segmentation and offers suggestions for improving their use in medical imaging tasks. The code is available at https://github.com/yunzhuC/MCSD.

O013 Sleep and overnight emotional memory in older adults with depression: A high-density EEG investigation

Abstract Older adults with depression have a high incidence of sleep disturbance which is posited to be mechanistically involved in maladaptive overnight emotional memory consolidation. In older adults (≥ 50 years) with and without depression, we aimed to compare group differences in overnight emotional memory and rapid eye movement (REM) and non-rapid eye movement (NREM) (N2 and N3) sleep disturbance. Secondly, we investigated the relationship between emotional memory consolidation, self-report emotional valence and arousal perception, and sleep disturbance. Participants underwent overnight PSG with high-density EEG. An emotional memory image task with concurrent subjective emotional arousal and valence rating was completed before and after sleep. REM sleep disturbance was measured by REM sleep duration, global REM gamma and alpha activity and REM EEG arousal index. NREM sleep disturbance was measured by NREM sleep duration, global NREM delta, alpha and sigma power, and NREM EEG arousal index. T-tests and non-parametric tests were used for group comparisons. Linear regressions were used to assess relationships between sleep disturbance and emotional memory. Twenty-two older adults (Depression: n = 12, Control: n = 10) with a mean age of 63.7 ± 6.5 years completed the study. Older adults with depression demonstrated differences in overnight perception of emotional valence and arousal for negative information, suggesting sleep may be involved in emotion perception. Global delta power in NREM was reduced in older adults with depression, suggestive of homeostatic alterations. However, no robust associations between overnight memory consolidation, emotional valence or arousal and REM or NREM sleep disturbance were observed.

Versatility Evaluation of Landslide Risk with Window Sizes and Sampling Techniques Based on Deep Learning

Across the globe, landslides cause significant loss of life, injuries, and widespread damage to homes and infrastructure. Therefore, assessing and analyzing landslide hazards is crucial to human, environmental, cultural, economic, and social sustainability. This study utilizes ArcGIS 10.8 and Python 3.9 to create landslide databases for Niigata Prefecture (NIG), Iwate and Miyagi Prefectures (IWT-MYG), and Hokkaido (HKD), drawing on data obtained from the National Research Institute for Earth Science and Disaster Resilience, Japan. A distinguishing feature of this study is the application of a Convolutional Neural Network (CNN), which significantly outperforms traditional machine learning models in image-based pattern recognition by extracting contextual information from surrounding areas, a distinct advantage in image and pattern recognition tasks. Unlike conventional methods that often require manual feature selection and engineering, CNNs automate feature extraction, enabling a more nuanced understanding of complex patterns. By experimenting with CNN input window sizes ranging from 3 × 3 to 27 × 27 pixels and employing diverse sampling techniques, we demonstrate that larger windows enhance the model’s predictive accuracy by capturing a wider range of environmental interactions critical for effective landslide modeling. CNN models with 19 × 19 pixel windows typically yield the best overall performance, with CNN-19 achieving an AUC of 0.950, 0.982 and 0.969 for NIG, HKD, and IWT-MYG, respectively. Furthermore, we improve prediction reliability using oversampling and a random window-moving method. For instance, in the NIG region, the AUC of the oversampling CNN-19 is 0.983, while the downsampling AUC is 0.950). These techniques, less commonly applied in traditional machine learning approaches to landslide detection, help address the issue of data imbalance often seen in landslide datasets, where instances of landslides are far outnumbered by non-landslide occurrences. While challenges remain in enhancing the model’s generalization, this research makes significant progress in developing more robust and adaptable tools for landslide prediction, which are vital for ensuring environmental and societal resilience.

Comprehensive image quality comparison of conventional and new flat panel detectors under bedside chest radiography beam conditions.

Recently, a novel wireless flat-panel detector with auto-exposure control has become available. This study aimed to elucidate the potential advantages of the new detector over conventional detectors through a comprehensive analysis of the physical image quality characteristics. Measurements were conducted on two models: new (720C) and conventional (710C) versions; this assessment was performed by assuming the beam quality for bedside chest radiography, utilizing a portable device for X-ray exposure. The detective quantum efficiency (DQE) was computed based on the presampled modulation transfer function (MTF) and normalized noise power spectrum. The validity of the DQE results was verified through the visualization of the analog blurring components and a detailed analysis of the noise components. The spatial frequency at which the presampled MTF value reached 10% was 5.2 cycles/mm for 720C and 3.9 cycles/mm for 710C. The full width at half-maximum of the spatial spreading of analog components was estimated at 0.09mm for 720C and 0.14mm for 710C by the visualization. Regarding the DQE, 720C was superior under low-dose conditions despite no significant differences being observed under high-dose conditions. The new detector demonstrated superior resolution characteristics compared with the conventional detector and an improvement in the DQE under low-dose conditions. However, similar to the conventional detector, a significant dose dependence caused by a structural factor was confirmed for the DQE. These results suggest the existence of an appropriate dose range for maximizing detector performance and provide insights crucial for optimization tasks in the X-ray imaging.

Data augmentation via warping transforms for modeling natural variability in the corneal endothelium enhances semi-supervised segmentation.

Image segmentation of the corneal endothelium with deep convolutional neural networks (CNN) is challenging due to the scarcity of expert-annotated data. This work proposes a data augmentation technique via warping to enhance the performance of semi-supervised training of CNNs for accurate segmentation. We use a unique augmentation process for images and masks involving keypoint extraction, Delaunay triangulation, local affine transformations, and mask refinement. This approach accurately captures the natural variability of the corneal endothelium, enriching the dataset with realistic and diverse images. The proposed method achieved an increase in the mean intersection over union (mIoU) and Dice coefficient (DC) metrics of 17.2% and 4.8% respectively, for the segmentation task in corneal endothelial images on multiple CNN architectures. Our data augmentation strategy successfully models the natural variability in corneal endothelial images, thereby enhancing the performance and generalization capabilities of semi-supervised CNNs in medical image cell segmentation tasks.

Functional localization of the human auditory and visual thalamus using a thalamic localizer functional magnetic resonance imaging task

Abstract Functional magnetic resonance imaging (fMRI) of the auditory and visual sensory systems of the human brain is an active area of investigation in the study of human health and disease. The medial geniculate nucleus (MGN) and lateral geniculate nucleus (LGN) are key thalamic nuclei involved in the processing and relay of auditory and visual information, respectively, and are the subject of blood-oxygen-level-dependent (BOLD) fMRI studies of neural activation and functional connectivity in human participants. However, localization of BOLD fMRI signal originating from neural activity in MGN and LGN remains a technical challenge, due, in part, to the poor definition of boundaries of these thalamic nuclei in standard T1-weighted and T2-weighted magnetic resonance imaging sequences. Here, we report the development and evaluation of an auditory and visual sensory thalamic localizer (TL) fMRI task that produces participant-specific functionally-defined regions of interest (fROIs) of both MGN and LGN, using 3 Tesla multiband fMRI and a clustered-sparse temporal acquisition sequence, in less than 16 minutes of scan time. We demonstrate the use of MGN and LGN fROIs obtained from the TL fMRI task in standard resting-state functional connectivity (RSFC) fMRI analyses in the same participants. In RSFC analyses, we validated the specificity of MGN and LGN fROIs for signals obtained from primary auditory and visual cortex, respectively, and benchmarked their performance against alternative atlas- and segmentation-based localization methods. The TL fMRI task and analysis code (written in Presentation and MATLAB, respectively) have been made freely available to the wider research community.

Novelty Classification Model Use in Reinforcement Learning for Cervical Cancer.

Cervical cancer significantly impacts global health, where early detection is piv- otal for improving patient outcomes. This study aims to enhance the accuracy of cervical cancer diagnosis by addressing class imbalance through a novel hybrid deep learning model. The proposed model, RL-CancerNet, integrates EfficientNetV2 and Vision Transformers (ViTs) within a Reinforcement Learning (RL) framework. EfficientNetV2 extracts local features from cervical cytology images to capture fine-grained details, while ViTs analyze these features to recognize global dependencies across image patches. To address class imbalance, an RL agent dynamically adjusts the focus towards minority classes, thus reducing the common bias towards majority classes in medical image classification. Additionally, a Supporter Module incorporating Conv3D and BiLSTM layers with an attention mechanism enhances contextual learning. RL-CancerNet was evaluated on the benchmark cervical cytology datasets Herlev and SipaKMeD, achieving an exceptional accuracy of 99.7%. This performance surpasses several state-of-the-art models, demonstrating the model's effectiveness in identifying subtle diagnostic features in complex backgrounds. The integration of CNNs, ViTs, and RL into RL-CancerNet significantly improves the diagnostic accuracy of cervical cancer screenings. This model not only advances the field of automated medical screening but also provides a scalable framework adaptable to other medical imaging tasks, potentially enhancing diagnostic processes across various medical domains.

Self-supervised learning for chest computed tomography: training strategies and effect on downstream applications.

Self-supervised pre-training can reduce the amount of labeled training data needed by pre-learning fundamental visual characteristics of the medical imaging data. We investigate several self-supervised training strategies for chest computed tomography exams and their effects on downstream applications. We benchmark five well-known self-supervision strategies (masked image region prediction, next slice prediction, rotation prediction, flip prediction, and denoising) on 15M chest computed tomography (CT) slices collected from four sites of the Mayo Clinic enterprise, United States. These models were evaluated for two downstream tasks on public datasets: pulmonary embolism (PE) detection (classification) and lung nodule segmentation. Image embeddings generated by these models were also evaluated for prediction of patient age, race, and gender to study inherent biases in models' understanding of chest CT exams. The use of pre-training weights especially masked region prediction-based weights, improved performance, and reduced computational effort needed for downstream tasks compared with task-specific state-of-the-art (SOTA) models. Performance improvement for PE detection was observed for training dataset sizes as large as with a maximum gain of 5% over SOTA. The segmentation model initialized with pre-training weights learned twice as fast as the randomly initialized model. While gender and age predictors built using self-supervised training weights showed no performance improvement over randomly initialized predictors, the race predictor experienced a 10% performance boost when using self-supervised training weights. We released self-supervised models and weights under an open-source academic license. These models can then be fine-tuned with limited task-specific annotated data for a variety of downstream imaging tasks, thus accelerating research in biomedical imaging informatics.

Deep learning-based temporal deconvolution for photon time-of-flight distribution retrieval.

The acquisition of the time of flight (ToF) of photons has found numerous applications in the biomedical field. Over the last decades, a few strategies have been proposed to deconvolve the temporal instrument response function (IRF) that distorts the experimental time-resolved data. However, these methods require burdensome computational strategies and regularization terms to mitigate noise contributions. Herein, we propose a deep learning model specifically to perform the deconvolution task in fluorescence lifetime imaging (FLI). The model is trained and validated with representative simulated FLI data with the goal of retrieving the true photon ToF distribution. Its performance and robustness are validated with well-controlled in vitro experiments using three time-resolved imaging modalities with markedly different temporal IRFs. The model aptitude is further established with in vivo preclinical investigation. Overall, these in vitro and in vivo validations demonstrate the flexibility and accuracy of deep learning model-based deconvolution in time-resolved FLI and diffuse optical imaging.

A Leaf Chlorophyll Content Estimation Method for Populus deltoides (Populus deltoides Marshall) Using Ensembled Feature Selection Framework and Unmanned Aerial Vehicle Hyperspectral Data

Leaf chlorophyll content (LCC) is a key indicator in representing the photosynthetic capacity of Populus deltoides (Populus deltoides Marshall). Unmanned aerial vehicle (UAV) hyperspectral imagery provides an effective approach for LCC estimation, but the issue of band redundancy significantly impacts model accuracy and computational efficiency. Commonly used single feature selection algorithms not only fail to balance computational efficiency with optimal set search but also struggle to combine different regression algorithms under dynamic set conditions. This study proposes an ensemble feature selection framework to enhance LCC estimation accuracy using UAV hyperspectral data. Firstly, the embedded algorithm was improved by introducing the SHapley Additive exPlanations (SHAP) algorithm into the ranking system. A dynamic ranking strategy was then employed to remove bands in steps of 10, with LCC models developed at each step to identify the initial band subset based on estimation accuracy. Finally, the wrapper algorithm was applied using the initial band subset to search for the optimal band subset and develop the corresponding model. Three regression algorithms including gradient boosting regression trees (GBRT), support vector regression (SVR), and gaussian process regression (GPR) were combined with this framework for LCC estimation. The results indicated that the GBRT-Optimal model developed using 28 bands achieved the best performance with R2 of 0.848, RMSE of 1.454 μg/cm2 and MAE of 1.121 μg/cm2. Compared with a model performance that used all bands as inputs, this optimal model reduced the RMSE value by 24.37%. In addition to estimating biophysical and biochemical parameters, this method is also applicable to other hyperspectral imaging tasks.

PIPET: A Pipeline to Generate PET Phantom Datasets for Reconstruction Based on Convolutional Neural Network Training

There has been a strong interest in using neural networks to solve several tasks in PET medical imaging. One of the main problems faced when using neural networks is the quality, quantity, and availability of data to train the algorithms. In order to address this issue, we have developed a pipeline that enables the generation of voxelized synthetic PET phantoms, simulates the acquisition of a PET scan, and reconstructs the image from the simulated data. In order to achieve these results, several pieces of software are used in the different steps of the pipeline. This pipeline solves the problem of generating diverse PET datasets and images of high quality for different types of phantoms and configurations. The data obtained from this pipeline can be used to train convolutional neural networks for PET reconstruction.

Long-term neuroplasticity in language networks after anterior temporal lobe resection.

Anterior temporal lobe resection (ATLR) is an effective treatment for drug-resistant temporal lobe epilepsy (TLE), although language deficits may occur after both left and right ATLR. Functional reorganization of the language network has been observed in the ipsilateral and contralateral hemispheres within 12 months after ATLR, but little is known of longer-term plasticity effects. Our aim was to examine the plasticity of language functions up to a decade after ATLR, in relation to cognitive profiles. We examined 24 TLE patients (12 left [LTLE]) and 10 controls across four time points: pre-surgery, 4 months, 12 months, and ~9 years post-ATLR. Participants underwent standard neuropsychological assessments (naming, phonemic, and categorical fluency tests) and a verbal fluency functional magnetic resonance imaging (fMRI) task. Using a flexible factorial design, we analyzed longitudinal fMRI activations from 12 months to ~9 years post-ATLR, relative to controls, with separate analyses for people with hippocampal sclerosis (HS). Change in cognitive profiles was correlated with the long-term change in fMRI activations to determine the "efficiency" of reorganized networks. LTLE patients had increased long-term engagement of the left extra-temporal and contralateral temporal regions, with better language performance linked to bilateral activation. Those with HS exhibited more widespread bilateral activations. RTLE patients showed plasticity in the left extra-temporal regions, with better language outcomes associated with these areas. Both groups of patients achieved cognitive stability over 9 years, with more than 50% of LTLE patients improving. Older age, longer epilepsy duration, and lower pre-operative cognitive reserve negatively affected long-term language performance. Neuroplasticity continues for up to ~9 years post-epilepsy surgery in LTLE and RTLE, with effective language recovery linked to bilateral engagement of temporal and extra-temporal regions. This adaptive reorganization is associated with improved cognitive outcomes, challenging the traditional view of localized surgery effects. These findings emphasize the need for early intervention, tailored pre-operative counseling, and the potential for continued cognitive gains with extended post-ATLR rehabilitation.