Tie “Thomas” Luo, Ph.D., associate professor in the University of Kentucky Department of Electrical and Computer Engineering, has received the Best Paper Award at the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) 2025 Workshop on Pattern Mining and Machine Learning for Bioinformatics for his recent work titled “Unlocking Neural Transparency: Jacobian Maps for Explainable AI in Alzheimer’s Detection.”
The study introduces a novel approach that tackles one of the most critical bottlenecks in AI-powered medical diagnostics: interpretability. While deep learning models have shown high accuracy in detecting Alzheimer’s disease, their “black box” nature has limited clinical adoption. Luo’s paper proposes a unique solution by using Jacobian Maps, a pre-model technique that captures localized structural changes in brain imaging data before training any neural network model.
“Our goal was to make deep learning not only accurate but also understandable and trustworthy to medical professionals,” said Luo. “Jacobian Maps act like a bridge between model predictions and known neuroanatomical biomarkers of Alzheimer’s.”
The method transforms MRI or CT brain scans into interpretable 3D maps that encode local volumetric changes, enabling the model to "see" patterns that correlate with disease progression. The Jacobian Maps-enhanced pipeline achieved state-of-the-art performance, with 95.2% accuracy in classifying cognitively normal patients and 90.2% for those with mild or moderate dementia. Additionally, it provided interpretable visual heatmaps that aligned well with clinical knowledge, particularly highlighting changes in the frontal-temporal region, a key biomarker area for Alzheimer’s.
In addition to the award-winning paper, Luo also presented a main conference paper titled “Enabling Heterogeneous Adversarial Transferability via Feature Permutation Attacks,” which introduced long-range dependencies, akin to global context, into convolutional neural networks, enabling them to mimic behaviors seen in vision transformers and multi-layer perceptrons for better generalization cross neural network architectures.
The 29th PAKDD was held in June in Sydney, Australia, attracting approximately 400 researchers from around the world. The main conference received over 700 submissions, of which only 168 papers were accepted.
