Abstracts

Rationale: Artificial intelligence (AI) is a promising approach to improve the detection of subtle lesions via structural neuroimaging in common forms of epilepsy such as temporal lobe epilepsy (TLE). However, AI tools (e.g., convolutional neural networks; CNNs) can have opaque reasoning (i.e., black-box), which hinders clinical adoption. Therefore, there is a great need for tools with both adequate predictive performance and clarity in the reasoning for decisions. Here we present a framework for a tool that can derive visual signatures of single-patient clinical features in a visually explorable way.

Methods: We used 1178 T1-weighted images (589 temporal lobe epilepsy; TLE, 589 healthy controls; HC) from 12 surgical centers to develop Semantic Exploration and Explainability using a Generative Adversarial Autoencoder Network (SEE-GAAN; Figure 1A) a generative AI framework for medical image exploration. SEE-GAAN first extracts a set of meaningful features from medical images (the ‘autoencoder’ part) in an unsupervised manner, which are then used to synthesize a sequence of images (the ‘generative adversarial’ part) that visualize the presentation of clinical features. SEE-GAAN was first used to explore the presentation of epilepsy (TLE vs HC; Figure 1B) and age (years) on T1-weighted images. SEE-GAN features were then used to train a support vector machine (SVM) to discriminate between TLE vs HC. A CNN (i.e., a supervised method) was then trained on the same task using the same images for comparison.

Results: Figure 1C displays the SEE-GAAN ‘global’ presentation of epilepsy on T1-weighted images (blue areas indicate greater signal, green areas indicate less signal). This epilepsy signal is centered on the limbic circuit, with subcortical structures like hippocampus and thalamus, along with lateral temporal neocortex and the insula. Figure 1D shows that at the single-subject level (i.e., the ‘local’ signal), the epilepsy pattern appears related to patient characteristics, in this case side of seizure onset. Next, we investigated age as a continuous feature, finding that for older persons, a signal developed associated with decreased grey matter. Figure 2 displays the subtraction of the signal between TLE and HC, demonstrating that at younger ages, there is no difference (i.e., the yellow color). However, as age increases, the aging signal becomes stronger in the patients. Finally, SVM and CNN accuracy for discriminating between TLE vs HC was 81.2% and 85.8%, respectively.


Conclusions: This proof of concept SEE-GAAN application shows: 1) An epilepsy structural signature can be detected using SEE-GAAN synthetic images at an individual patient level, 2) the SEE-GAAN sequence for age matches previous findings of accelerated aging in epilepsy, and 3) SEE-GAAN features achieved lower but generally comparable accuracy relative to black box methods like CNN. SEE-GAAN sequences facilitate exploratory medical image analysis in neurological disease and offer ways to explore multiple clinical and demographic features simultaneously at an individual patient level.

Funding: K01NS124831