Register for free to
view the full innovation profile.
whole-slide cartography is the segmentation by tissue type: tumor, stroma, mucosa,.. We additionally speed up the computation by 40% with smart preprocessing
About
We digitized and annotated a dataset of >150 whole-slide images of H&E stained colon resections. The entire foreground is annotated into 7 classes: active tumor cells, mucosa, muscle tissue, fat, necrosis, inflammation, mucus. We then trained various network architectures (CNNs) on our on-site GPU-deep-learning cluster to determine the most suited network. Instead of simply tiling the whole-slide-image into non-overlapping patches and labeling each patch, we first segment the foreground into superpixels. A random subset of tiles per superpixel is then selected and classified and the overall superpixel label is inferred. This not only yields a significant speed-up, since not all tiles are classified, but it even increases the overall accuracy slightly. The resulting “tissue map” can be overlaid transparently on top of the H&E slide to give the pathologist a good overview. The absence or presence of areas (even small ones) of the “active tumor cells” class indicates for each slide whether it is healthy or not. Additionally, the tissue composition in the tumor is a prognostic factor: more stroma indicates better prognosis. It is also straightforward to derive the tumor invasion front from the tumor area. The shape of the invasion front indicates the tumor’s growth pattern. By running further post-processing it should furthermore be possible to derive the pT classification (ongoing work). We have also digitized the dataset with scanners from five vendors and are currently working on making the classification step robust to variances in color. Other work in the field of GI pathology: we have also trained an AI-based tumor-budding classifier for IHC (AE1/AE3 “pan-cytokeratin”) stains of the same dataset. We have additionally created a multi-centric dataset of gastroesophageal tumors comprising H&E and 9 other IHC stains and are currently developing a co-expression analysis workflow.
Key Benefits
whole-slide cartography provides basis for various medical applications - tumor absence/presence - invasion depth (pT in TNM classification) - invasion front detection (e.g. detect tumor buds here; look at tumor microenvironment; classify tumor growth pattern) - tumor composition (high stroma indicates better prognosis) - analyze only tumor tissue w.r.t. typical gene mutations, e.g. MSI in case of colorectal cancer Our solution yields very accurate classification accuracy (>95%) and is at the same time 40% faster compared to the standard approach, where the WSI is separated into (non-overlapping) tiles and each tile is classified. Our solution is available as a stand-alone C++ library that can be integrated into 3rd party desktop or cloud-based solutions. Additionally, we have integrated it into our own (re-brandable) advanced desktop microscopy whole-slide viewer that includes an "App-center" for analysis apps. Computational Pathology is growing field, motivated on the clinical side by - lack of pathologists on the one hand - increasing demand (precision/personalized medicine) on the other hand In the context of pharmacological research, the main motivation is to streamline quantification through automation and remove any human bias inherently present when a pathologist manually evaluates whole-slide.
Applications
research: - quantitative evaluation ("whole-slide to Excel") - automation / batch analysis - removal of human bias clinical: - QC ("2nd look") - automation (computer assisted diagnosis) - more accurate risk stratification (computer can calculate far more complex scores compare to human)