Multi-Vector SIVQ as a Tool for Autonomous Tingible Body Macrophage Localization
Jennifer A Hipp, Gaurav Sharma, Jason D Hipp, Jerome Cheng, Ulysses J Balis, Megan Lim, Kojo S Elenitoba-Johnson. University of Michigan, Ann Arbor
Background: The recent availability of digital whole slide data sets has created new opportunities for pathologists to perform numerical and quantitative assessment of histologic features. Observing the above reality, the transition from pathologist-dependent analogue analysis of histopathologic material to a pathologist-enabling and computationally assisted interpretation paradigm underscores the need for new algorithms capable of recognition of discriminative features of different histologic entities. Spatially Invariant Vector Quantization (SIVQ) is a feature-recognition algorithm which exploits the continuous symetry of ring vectors within textural domains of any image. Follicular Hyperplasia (FH) is a benign reactive condition characterized by variably sized and enlarged follicles with presence of Tingible body macrophages (TBM) in the germinal centers (GC). Using a prototype SIVQ-based routine, we investigated the feasibility of digital recognition of TBM as a computational aid for FH with the aim of developing it as a discriminant from follicular lymphoma (FL).
Design: Three SIVQ ring vectors (feathery cytoplasm, prominent nucleoli and apoptotic bodies) were defined to identify TBM at intermediate power (20x). Ten cases of FH were selected from whole-slide-image of tissue microarray prepared from previously diagnosed cases. In each case, four GC rich areas were selected for SIVQ analysis by two observers.Inclusion criteria included presence of a GC and an unequivocal histopathological diagnosis. Performance of SIVQ analysis was compared with that of manual review of same areas.
Results: After carrying out the aforementioned event qualification tabulation, sensitivities and specificities for vector 1, 2 and three were (57% and 75%), (71% and 68%) and (94% and 65%), respectively. Overall, vector 1 identified 29/48 TBMs, vector 2 identified 32/48 TBMs, while Vector 3 identified 45/48 TBMs. With taken in aggregate, the visual estimate of all three sets of vector events confirmed a synergistic effect for both elevating sensitivity and specificity for TBM detection.
Conclusions: Multi-vector based SIVQ successfully recognized TBMs in these FH cases with over 94% accuracy. In broader terms, our results indicate that pathologist-educated image analysis algorithms are enabling adjunct tools for automated recognition of distinct pathological entities. It is anticipated that image-dependent feature-recognition algorithms will emerge as transformative tools in implementation of digital histopathological analysis workflows.
Wednesday, March 21, 2012 1:00 PM
Poster Session VI # 243, Wednesday Afternoon