Digitized pathology: theory and experiences in automated tissue-based virtual diagnosis

Abstract

Aims: To describe the theory and develop an automated virtual slide screening system. Theoretical considerations. Tissue-based diagnosis separates into (a) sampling procedure to allocate the slide area containing diagnostic information, and (b) evaluation of diagnosis from the selected area. Nyquist's theorem broadly applied in acoustics, serves to presetting the sampling accuracy. Tissue-based diagnosis relies on two different information systems: (a) texture, and (b) object information. Texture information can be derived by recursive formulas without image segmentation. Object information requires image segmentation and feature extraction. Both algorithms complete another to a "self-learning" classification system.

Methods: Non-overlapping compartments of the original virtual slide (image) are chosen at random with predefined error-rate (Nyquist's theorem). The standardized image compartments are subject for texture and object analysis. The recursive formula of texture analysis computes median gray values and local noise distribution. Object analysis includes automated measurements of immunohistochemically stained slides. The computations performed at different magnifications (x 2, x 4.5, x 10, x 20, x 40) are subject to multivariate statistically analysis and diagnosis classification.

Results: A total of 808 lung cancer cases of diagnoses groups: cohort (1) normal lung (318 cases) - cancer (490 cases); cancer subdivided: cohort (2) small cell lung cancer (10 cases) - non-small cell lung cancer (480 cases); non-small cell lung cancer subdivided: cohort (3) squamous cell carcinoma (318 cases) - adenocarcinoma (194 cases) - large cell carcinoma (70 cases) was analyzed. Cohorts (1) and (2) were classified correctly in 100%, cohort (3) in more than 95%. The selected area can be limited to 10% of the original image without increased error rate. A second approach included 233 breast tissue cases (105 normal, 128 breast carcinomas) and 88 lung tissue cases (58 normal, 38 cancer). Texture analysis revealed a correct classification with only 10 training set cases in >92% for both, breast and lung tissue.

Conclusions: The developed system is a fast and reliable procedure to fulfill all requirements for an automated "pre-screening" of virtual slides in tissue-based diagnosis.