Organ Proximity Analysis: A Novel Approach to Spleen Localization for Accurate Injury Grading in Abdominal CT Scans

Accurate grading of splenic injuries in abdominal CT scans is paramount for diagnosing and treating abdominal trauma. However, automating the segmentation of an injured spleen is a formidable task, given the numerous anomalies like shape, location, texture, and size variations. Deep neural networks often require extensive and expensive annotated datasets to achieve precise spleen segmentation. This paper presents a novel approach that integrates segmentation and localization techniques for spleen analysis. Initially, we segment the spleen using a segmentation model trained on limited normal spleen data. We then employ a localization decision-making process to determine whether the segmented spleen exhibits an out-of-distribution shape and location. If not, we utilize the segmentation map for cropping. In cases where the spleen displays atypical features, we leverage information from nearby organs to accurately locate the spleen. The cropped spleen is then processed through a 3D DenseNet, a deep convolutional network, for spleen classification, categorizing the injuries into normal, low (grades I to III) and high (grades IV & V) grade injuries following the American association for the surgery of trauma injury scoring scale. This innovative approach not only achieves accuracy levels comparable to segmentation-based methods but also reduces dependence on extensive annotated datasets, making it a noteworthy contribution to abdominal trauma diagnosis and treatment. Code https://github.com/hamghalam/Proximity-Analysis.