Computer aided diagnosis of acute intracranial hemorrhage on brain CT

Abstract

Acute intracranial hemorrhage (AIH) is a major cause of neurological disturbance or complication of head injury. Its presence dictates different management strategy. In modern medicine, detection of AIH relies on the use of brain computed tomography (CT). But diagnosis of AIH can become difficult when the lesion is inconspicuous or the reader is inexperienced. The objective of the current project is to develop a computer aided diagnosis (CAD) system that improves the diagnostic performance of AIH on CT by clinicians. A total of 186 cases, including all 62 continuous cases that showed AIH not more than 1cm in size obtained during a 6 month period, and 124 randomly selected controls that were obtained during the same period, were retrospectively collected from the CT archive of Princess Margaret Hospital. The imaging diagnoses were established by consensus of two experienced radiologists. A CAD was designed and implemented. It reads and processes standard DICOM image files. Intracranial contents are segmented from the CT images, which are then subjected to denoising and adjustment for CT cupping artifacts. AIH candidates are extracted from the intracranial contents based on top-hat transformation and subtraction between two sides of the image about the mid-sagittal plane. AIH candidates are registered against a normalized coordinate system such that the candidates are rendered anatomical information. True AIH is differentiated from mimicking normal variants or artifacts by a knowledge based classification system incorporating rules that make use of quantified imaging features and anatomical information. The CAD algorithm was manually trained using 40 positive and 80 control cases. In the validation test using the remaining 22 positive and 44 control cases, the system achieved sensitivity of 83% for small (< 1cm) AIH lesions on a per lesion basis or 100% on a per case basis, and false positive rate 0.020 per image or 0.29 per case. In the observer performance study, 7 emergency physicians, 7 radiology residents and 6 radiology specialists were recruited as readers of 60 sets of brain CT selected from the same 186 case collection, including 30 positive cases and 30 controls. Each reader read the same 60 cases twice, first without, then with the prompts produced by the CAD system. The clinicians rated their confidence in diagnosing each case of showing AIH in both reading modes. The results were analyzed using the multiple-reader, multiple-case receiver operating characteristic (MRMC ROC) paradigm, which showed significantly improved performance for emergency physicians, average area under the ROC curve (Az) significantly increased from 0.83 to 0.95 without and with the support of CAD. Az for radiology residents and specialists also improved, from 0.94 to 0.98 and from 0.97 to 0.98 respectively.. In summary, a CAD system which boasted high sensitivity and low false positive rate has been developed. MRMC ROC study confirmed that it can improve diagnostic performance of clinicians, especially emergency physicians. It is anticipated that such a system can reduce diagnostic errors and improve patient care when it is integrated in the clinical environment for daily operation.