Assessing the performance of four different categories of histological criteria in brain tumours grading by means of a computer‐aided diagnosis image analysis system

• Published in: Journal of microscopy (Oxford) Vol. 260; no. 1; pp. 37 - 46
• Main Authors: KOSTOPOULOS, S., KONSTANDINOU, C., SIDIROPOULOS, K., RAVAZOULA, P., KALATZIS, I., ASVESTAS, P., CAVOURAS, D., GLOTSOS, D.
• Format: Journal Article
• Language: English
• Published: England Wiley Subscription Services, Inc 01-10-2015
• Subjects: Algorithms; Brain Neoplasms - diagnosis; Brain Neoplasms - pathology; Brain Neoplasms - ultrastructure; Brain tumours; classification; computer‐aided diagnosis; Diagnostic Errors - prevention & control; graphics processing unit; Greece; Histological Techniques; histopathology; Humans; Image Processing, Computer-Assisted - methods; Image Processing, Computer-Assisted - standards; microscopy; Neoplasm Grading; Greece; graphics processing unit; histopathology; microscopy; Brain tumours; classification; computer-aided diagnosis
• ISSN: 0022-2720; 1365-2818
• DOI: 10.1111/jmi.12264

Summary
Brain tumours are considered one of the most lethal and difficult to treat forms of cancer, with unknown aetiology and lack of any realistic screening. In this study, we examine, whether the combination of descriptive criteria, used by expert histopathologists in assessing histologic tissue samples, and quantitative image analysis features may improve the diagnostic accuracy of brain tumour grading. Data comprised 61 cases of brain cancers (astrocytomas, oligodendrogliomas, meningiomas) collected from the archives of the University Hospital of Patras, Greece. Incorporating physician's descriptive criteria and image analysis's quantitative features into a discriminant function, a computer‐aided diagnosis system was designed for discriminating low‐grade from high‐grade brain tumours. Physician's descriptive features, when solely used in the system, proved of high discrimination accuracy (93.4%). When verbal descriptive features were combined with quantitative image analysis features in the system, discrimination accuracy improved to 98.4%. The generalization of the proposed system to unseen data converged to an overall prediction accuracy of 86.7% ± 5.4%. Considering that histological grading affects treatment selection and diagnostic errors may be notable in clinical practice, the utilization of the proposed system may safeguard against diagnostic misinterpretations in every day clinical practice.
Lay Description
Brain tumours are considered one of the most lethal and difficult to treat forms of cancer, with unknown aetiology and lack of any realistic screening. Even though neuroimaging combined with tissue microscopy are fundamentally important for patient management, the potential of diagnostic errors still remains substantially high. One of the approaches that have been proposed, as a precautionary measure against potential diagnostic misinterpretations, is the utilization of computer‐aided diagnostic systems as second opinion tools.
In this study, we have designed computer software for computer‐aided diagnosis of brain tumours. The software functions by (i) evaluating image properties, after suitable processing of the brain tumour's microscopy images, (ii) quantifying the physician's descriptive criteria, which are verbal expressions used by the physician when assessing the tumour's tissue samples under the microscope and (iii) characterizing the type of brain tumour by means of a specially designed classification engine that is fed with information extracted from the tumour's imaging properties and the physician's descriptive criteria. For the design of the software, we have used biopsy material of verified brain cancer cases (such as astrocytomas, oligodendrogliomas, meningiomas) from the archives of the University Hospital of Patras, Greece. Specially prepared specimens were provided for examination under the microscope that had been stained with Hematoxylin and Eosin dye. The microscope was connected to a digital camera and images were stored into a computer that hosted the software. Five nonoverlapping digital images were, thus, collected from each tumour and a number of properties were evaluated from the texture, shape and distribution of cell‐nuclei contained within each microscopy image; cell‐nuclei have long been recognized as significant indicators of tumours malignancy. Additionally, eight descriptive features, as evaluated by the histopathologist, were quantified for incorporation in the developed software. Features included cellularity, mitoses, apoptosis, multinucleated, giant, vascular proliferation, necrosis and pleomorphism.
The diagnostic precision of the software in characterizing correctly tumours that were used for its design was perfect, however, when the system was presented to new, ‘unseen’ by the system verified data, the diagnostic precision dropped to 87%. That is a good estimate as to how such software would perform in a clinical environment. Considering that histological grading affects treatment selection and that diagnostic errors may be notable in clinical practice, the utilization of the proposed system may safeguard against diagnostic misinterpretations in every day clinical practice, especially by nonexperienced physicians.