Predicting Medial Temporal Lobe Integrity and Memory from Thalamic Microstructure Using Radiomics
Poster No:
1080
Submission Type:
Abstract Submission
Authors:
Rowan Lawrence1, Daniela Montaldi1, Alex Kafkas1
Institutions:
1University of Manchester, Manchester, Greater Manchester
First Author:
Co-Author(s):
Introduction:
Dual-process episodic memory models posit separation of recall (recognition with additional retrieval) and familiarity (simple recognition without further recall) in the medial temporal lobes (MTL). The MTL is functionally connected to other structures including nuclei of the thalamus (Aggleton et al., 2023), together supporting episodic memory. FMRI evidence suggests anterior nuclei support recall, while the medial / mediodorsal nuclei support familiarity (Kafkas et al., 2019). Radiomics, a recently-developed image analysis method, posits that images such as MRI contain microstructure and disease process information which can be extracted by mathematical transforms, providing high-dimensional datasets of quantitative features (Lambin et al., 2017). Features include shape, such as elongation and sphericity; features derived from the voxel intensity histogram; and higher-order features which quantify texture, reflecting spatial relationships of voxels in a region (Li et al., 2023). We used radiomics to establish relationships between thalamic integrity, episodic memory, and MTL damage in a large sample of patients with selective memory deficit.
Methods:
T1 images (0.83mm isotropic) were used to produce volumes of MTL structures using stereology, following a custom protocol based upon histological evidence (Insausti et al.,1998; Kivisaari et al., 2020). Patient volumes were compared to age-matched controls. Resulting Z-scores categorised patients as either hippocampal (HC N = 8), MTL cortical (MTLc N = 19), or both (MTL+ N = 43). T1 images were processed with FreeSurfer 7.2.0, and thalamic nuclei segmented with segmentThalamicNuclei (Iglesias et al., 2018). In-house Python scripts extracted radiomics features with the PyRadiomics library (van Griethuysen et al., 2017). Feature selection for each outcome variable was conducted with LASSO regression, with lambda hyperparameter tuned using leave-one-out cross-validation (LOOCV). Model fit was assessed with assess.glmnet (glmnet 4.7-1).
·Transverse slice through T1 of a Control subject showing thalamic nuclei ROIs following segmentation via FreeSurfer. Masks were collapsed into their respective nuclei groups before feature extraction.
Results:
Patients did not differ on tests of executive (Brixton; H(3) = 0.84, p = .840), semantic (Pyramids & Palm Trees; H(3) = 4.87, p = .182), or visuospatial (VOSP; H(3) = 2.95, p = .400) function, suggesting selective memory deficit. LASSO identified 35 features with non-zero coefficients which classified controls from patients, with accuracy of 87% (Fig. 2). A model fit to classify patient subtype (HC, MTLc, MTL+) showed prediction accuracy of 51%. 9 features with non-zero coefficients were related to recall memory. The model was overall significant F(9, 141) = 3.89, p <.001, Adj. R2 = .148 - two anterior thalamic features were significantly predictive of recall (both p <.05). 11 features were related to familiarity memory. The model was overall significant F(11, 101) = 6.53, p <.001, Adj. R2 = .352. An anterior thalamic texture feature was significantly predictive of familiarity (p = .040). 12 features were identified related to hippocampal volume, with a significant linear regression model F(12, 138) = 6.61, p <.001, Adj. R2 = .310. Multiple anterior and medial thalamic nuclei features were individually significant, including sphericity, elongation, size zone nonuniformity, and coarseness (anterior); and surface-volume ratio and elongation (medial).
·ROC of Control - Patient binomial classification using first-order and textural thalamic radiomics features
Conclusions:
We used radiomics-derived thalamic shape and texture information to characterise relationships between thalamus integrity, recall, familiarity, and upstream hippocampal atrophy in a dual-process episodic memory context. Despite primarily MTL damage and normal cognitive profile, patients could be classified based upon thalamic features suggesting downstream microstructural changes, which were additionally related to episodic memory performance. Future work will investigate microstructural bases of features using diffusion imaging, and attempt to improve subgroup classification with the use of image filters prior to feature extraction.
Learning and Memory:
Long-Term Memory (Episodic and Semantic) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Segmentation and Parcellation
Keywords:
Cognition
Learning
Limbic Systems
Memory
Modeling
Thalamus
Other - Radiomics
1|2Indicates the priority used for review
Provide references using author date format
Griethuysen, J.J.M. (2017), "Computational radiomics system to decode the radiographic phenotype", Cancer Research, vol. 77, no. 21, pp. 104-107
Iglesias, J.E. (2018), "A probabilistic atlas of the human thalamic nuceli combining ex vivo MRI and histology", NeuroImage, vol. 183, pp. 314-326
Insausti, R. (1998), "MR volumetric analysis of the human entorhinal, perirhinal, and temporopolar cortices", American Journal of Neuroradiology, vol. 189, pp. 659-671
Kafkas, A. (2020), "Thalamic-medial temporal lobe connectivity underpins familiarity memory", Cerebral Cortex, vol. 30, no. 6, pp. 3827-3837
Kivisaari, S.L. (2020), "The Perirhinal, Entorhinal, and Parahippocampal Cortices and Hippocampus: An Overview of Functional Anatomy and Protocol for Their Segmentation in MR Images", fMRI - Basics and Clinical Applications (eds.), Springer-Verlag, Berlin Heidelberg
Lambin, P. (2017), "Radiomics: the bridge between medical imaging and personalized medicine", Nature Reviews Clinical Oncology, vol. 6, no. 5, pp. 749-762
Li, M. (2023), "The radiomics features of the temporal lobe region related to menopause based on MR-T2WI can be used as potential biomarkers for AD", Cerebral Cortex, vol. 33, no. 14, pp. 9067-9078