Automated Detection of Histological Hallmarks in Frontotemporal Lobar Degeneration Using Deep Learning

Abstract

Frontotemporal lobar degeneration (FTLD) is a progressive neurodegenerative disease marked by distinct histological hallmarks, including Pick bodies. Manual identification is time-consuming, subjective, and requires expert neuropathologists. This study developed a convolutional neural network (CNN) for the automated detection of Pick bodies in histological images of FTLD. The model achieved 86.3% accuracy, 89.0% recall, and 0.91 ROC AUC, demonstrating its potential for objective and scalable identification of FTLD-related histopathological features, with applications for clinical diagnosis. Inference time per image was 0.042 seconds. Pixel density analysis revealed a significant difference between positive (mean 59.8) and negative (mean 47.3) regions. These findings support the feasibility of deep learning in neuropathology, enabling objective and scalable identification of FTLD-related changes. This approach offers potential for clinical integration, accelerated diagnosis, and expansion to other neurodegenerative disorders.