Deep learning-driven macroscopic AI segmentation model for brain tumor detection via digital pathology: Foundations for terahertz imaging-based AI diagnostics

Abstract

We used deep learning methods to develop an AI model capable of autonomously delineating cancerous regions in digital pathology images (H&E-stained images). By using a transgenic brain tumor model derived from the TS13-64 brain tumor cell line, we digitized a total of 187 H&E-stained images and annotated the cancerous regions in these images to compile a dataset. A deep learning approach was executed through DEEP:PHI, which abstracts Python coding complexities, thereby simplifying the execution of AI training protocols for users. By employing the Image Crop with Mask technique and patch generation method, we not only maintained an appropriate data class balance but also overcame the challenge of limited computing resources. This approach enabled us to successfully develop an AI training model that autonomously segments cancerous areas. This AI model enables the provision of guiding images for determining cancerous areas with minimal assistance from neuropathologists. In addition, the high-quality, large dataset curated for training using the proposed approach contributes to the development of novel terahertz imaging-based AI cancer diagnosis technologies and accelerates technological advancements.