Specialized Non-local Blocks for Recognizing Tumors on Computed Tomography Snapshots of Human Lungs

Abstract

This research endeavor is dedicated to the integration of specialized attentional mechanisms within the intricate web of deep neural network architectures aimed at discerning indications of lung carcinoma from monochromatic snapshots derived from computerized axial tomography. Within this exploration, we propose a myriad of adaptations to the traditional non-local blocks, infusing them with bespoke attentional nuances to resonate with the idiosyncrasies of medical imaging data. These bespoke adaptations ushered in discernible ameliorations in the performance metrics of the fundamental deep neural network model. Our solution facilitated a reduction in the model parameter count without compromising classification efficiency significantly. Additionally, it enabled a streamlined approach to feature extraction, contributing to enhanced interpretability and efficiency in the recognition process. These advancements were meticulously validated across test subsets meticulously curated from the Open Joint Monochrome Lungs Computer Tomography dataset, the Lung Image Database Consortium and Image Database Resource Initiative dataset, the Iraq-Oncology Teaching Hospital / National Center for Cancer Diseases dataset, Radiology Moscow and The Cancer Imaging Archive and from several others.