Privacy-Preserving Synthetic Histopathological Single-to-Multimodal Data Generation from Brain MRI Using Transfer Learning

Brain cancer detection is dependent on multiple diagnostic techniques. Histopathological diagnosis, although the most effective, requires the extraction of cancer cells, which is very risky and painful for a patient. Another popular noninvasive image-based diagnosis technique is magnetic resonance imaging (MRI). Brain diagnosis data based on MRI scans are highly sensitive and private. This study proposes a single-to-multimodal transformation technique that generates synthetic histopathological data from expert-labelled brain MRI datasets using transfer learning techniques. Furthermore, to preserve a patient’s privacy, an encryption module is used to encrypt the MRI image data and the respective histopathological notations. The Kruskal–Wallis statistical test is also used to analyze the radiogemomics dataset. The trained module is also encrypted, only to be accessed by authorized medical personnel. The transfer learning modules (CNN-based deep learning model, ViT, Resnet101, and YOLOv8) are used here and achieved 99.60% accuracy.