Semi-automated Classification of CT Scans in Traumatic Brain Injury Patients

Abstract

A 'silent epidemic' affecting millions worldwide every year is the Traumatic Brain Injury. Management of these patients essentially involves neuroimaging and noncontrast Computed Tomography (CT) scans are the first choice amongst doctors. However, interobserver variability, considered 'Achilles heel' amongst radiologists, can lead to missed diagnoses and grave consequences. This paper presents a hybrid approach for semi-automated classification of CT features according to Marshall CT Scheme. The proposed method uses template matching, artificial neural networks and active contours for segmentation of significant anatomical landmarks and estimation of haematoma volume on brain CT scans. The proposed method is efficient and robust in segmenting cross-sectional, noncontrast CT scans and has been evaluated on images from subjects with different ages and both genders. The hybrid method has an average ICC ≥ 0.97 and Jaccard Index ≥ 0.86 compared with the manual demarcations by radiology experts and performs better than the state of the art. Hence, the approach can be used to provide second opinions very close to the experts' intuition.

Automatic Quantification of Computed Tomography Features in Acute Traumatic Brain Injury

“…Multiple semiquantitative methods could be used, which can be an additional source of variability. 30…”

Automated Detection and Screening of Traumatic Brain Injury (TBI) Using Computed Tomography Images: A Comprehensive Review and Future Perspectives

“…The hematoma segmentation in CT imagery can be realized using rule-based models [ 23 , 41 , 76 ] or machine learning models [ 18 , 27 , 57 , 58 , 69 , 77 ]. Chan [ 23 ] developed a knowledge-based classification system from symmetry analysis to detect acute hematoma.…”

“…Muschelli et al [ 18 ] applied a voxel selection method based on handcrafted intensity featuring to detect ICH. Qureshi et al [ 76 ] have tried a semi-automated approach using ANN to perform initial pixel-wise categorization with an active contour for subsequent segmentation. Yao et al [ 59 ] generated super-pixels using the simple linear iterative clustering (SLIC) algorithm, and extracted statistical and textural features to automate hematoma segmentation.…”

“…It is evident from Table 3 and Table 4 that, in the existing studies for detection of hematoma regions, the widely used supervised classifiers are: the support vector machine (SVM) [ 42 , 51 , 71 ], random forest (RF) [ 18 , 57 , 69 ], artificial neural network (ANN) [ 41 , 76 ], probabilistic neural network (PNN) [ 70 ], Bayesian [ 56 ], multinomial logistic regression [ 67 ], tree bagger [ 58 ], and C4.5 [ 73 ].…”