Abstract: Artificial intelligence-based deep learning imaging frameworks are increasingly being developed to analyze large-scale, high-dimensional, and complex imaging data across scientific, biomedical, and clinical domains. Although the utilities have spread across domains and the implications are quite effective, conventional deep learning approaches rely heavily on manual annotation by experts to perform such tasks. This dependence on expert annotations is not only costly and time-consuming but also tends to bias models toward specific tasks, thereby limiting their internal representational capabilities. To mitigate these limitations, unsupervised and self-supervised learning approaches provide an alternative paradigm by enabling models to learn meaningful latent representations directly from the data without requiring extensive manual supervision.

In the context of brain imaging, these challenges are particularly significant. Brain MRI and histological brain images are inherently high-dimensional and exhibit a high degree of anatomical, pathological, and acquisition-related variability. The annotation of such images often requires expert clinical or neuroanatomical knowledge, making large-scale labeled datasets difficult to obtain. In addition, different brain image analysis tasks, such as disease detection, anatomical region classification, artifact restoration, and disease progression modeling, require representations that capture different levels of structural information.

The central focus of this work is to investigate how unsupervised and self-supervised models can learn, organize, and utilize latent representations for brain image analysis, disease diagnosis, and disease prognosis. More specifically, the work studies latent representations through the properties of compactness, separability, controllability, and disentanglement, depending on the complexity of the task being addressed.

This work presents a unified representational perspective for unsupervised brain image analysis across diagnosis, histological understanding, and prognosis. It demonstrates that as the task complexity increases, the latent space must satisfy progressively stronger structural and functional requirements. By developing models that better organize, control, and align latent representations, this work establishes that meaningful latent spaces can serve as the basis for robust and scalable unsupervised learning in brain imaging.

Event Details
Title: Unsupervised Modeling of Brain Images for Analysis, Diagnosis, and Prognosis of Neurological Diseases (PhD viva voce)
Date: June 30, 2026 at 02:30 PM
Venue: Google Meet (https://meet.google.com/ceh-uzzm-zsm)
Speaker: Ms. Ayantika Das (EE19D422)
Guide: Dr. Mohanasankar Sivaprakasam
Type: PHD seminar

Updated: